# Axiom public lab evidence report

## Pentesting & lab evidence

Detailed public report | Prepared 17 November 2026

A source-based review of the Axiom Windows laboratory records from 13-16 November 2026, with later interface screenshots. This document explains the testing method, requested file operations, recorded outcomes, changes between builds, and limits of the evidence. No new security test was executed to prepare it.

| Verified from saved event data | Count |
| --- | --- |
| Event records | 4,044 |
| Explicit allow/block decisions | 3,999 |
| GUARD_BLOCKED / GUARD_ALLOWED | 3,189 / 810 |
| Processing failures / other events | 25 / 20 |
| Latest complete saved standard sweep | 150 matches: 130 BLOCK + 20 ALLOW |

### What the results support

The saved logs support repeated tool-hook refusals and successful benign controls. Contemporaneous notes document earlier gaps and subsequent fixes. They do not establish that every adapter was exercised in a real authenticated agent conversation, that all named files physically existed, or that Axiom independently enforced every action at the operating-system boundary.

### Material exceptions

The first baseline has nine unknown outcomes. Sweep 067 is incomplete and the referenced sweep 073 is absent. Later sweep headers retain a stale build string. Historical direct malware execution did not produce new Axiom/Blackbox findings. These limits are retained in the report and downloadable data.

Reading guide: method and counting; harnesses and cases; repeated sweeps and hardening; file/resource and canary inventory; sample observations; OS boundary; screenshots; privacy and source index. Full row-level appendices are supplied as CSV files.

## 01 / How testing was performed

The documented environment was a Windows 11 VirtualBox laboratory with synthetic canary data. Builds changed throughout the exercise. Early 0.51 provisioning was a partial per-user install without service registration; an older 0.30 autostart remained. Later notes record installed service/desktop builds. Generated-source tests, partial installs, and installed-build checks must be read separately.

Sources: harness-inventory-2026-11-15.md:170-186; axiom-051-end-to-end-2026-11-15.md; pentest-2026-11-16-0.75.md:1-19.

| Stage | Recorded procedure and interpretation |
| --- | --- |
| Establish state | Identify build, installed client, hook contract, enabled policy and cached knowledge state. A displayed version is a snapshot, not a stamp on all accumulated history. |
| Construct input | Supply a hook-shaped event naming a tool and its arguments. File targets, write content and command descriptions are synthetic probe inputs. |
| Invoke boundary | Replay the provisioned hook/bridge, then interpret its host-specific response: JSON denial, cancellation flag, permission decision, or exit status. |
| Compare expectation | Sensitive test cases expect BLOCK; benign controls expect ALLOW. Preserve UNKNOWN or runner errors instead of treating them as permissions. |
| Fix and repeat | Document a gap, patch a rule/helper or adapter, repeat the targeted probes, and check the installed build where recorded. |
| Separate actual execution | Use distinct filesystem/process/network observations for real execution and OS enforcement. A refused hook input alone does not prove a payload ran or an OS call was denied. |

### Adapter contracts matter

The notes cover path keys such as file_path and absolute_path, nested tool arguments, lifecycle dispatch, and differing deny responses. The apparent 17-case tamper regression was traced to runner interpretation; the corrected runner reported eight denied tamper probes per hook across ten hooks. Source: pentest-2026-11-16-0.65.md:19-29.

Reproduction requires the matching build, policy, adapter and original controlled inputs. The public bundle supplies descriptions and sanitized event rows, not secrets, executable malware, exploit scripts, or a complete runnable reproduction kit.

## 02 / Reconcile the event streams

The vm-agent-evidence.jsonl snapshot contains 4,044 valid JSON records, from 2026-11-13 23:20:15 UTC to 2026-11-16 13:09:16 UTC. Both raw streams parsed without invalid lines. The public event CSV retains original source line numbers and categorical fields, but omits payloads and identifiers.

| Event type | Records |
| --- | --- |
| GUARD_ALLOWED | 810 |
| GUARD_BLOCKED | 3189 |
| GUARD_FAILED | 25 |
| HOOK_EVENT | 16 |
| SESSION_STARTED | 1 |
| PROMPT_SCANNED | 1 |
| POST_TOOL | 1 |
| STOPPED | 1 |

3,189 + 810 = 3,999 explicit allow/block decisions. The remaining 45 records are 25 processing failures and 20 lifecycle/hook events. The failures contain 15 JSONDecodeError and 10 ImportError codes. They must not be counted as successful blocks or benign allows.

### A separate audit stream

vm-audit.jsonl contains 1,368 records: 1,349 pre_tool_call, 16 pre_api_request and three lifecycle records. Explicit audit decisions are 532 block and 24 flag; 812 have no recorded decision. There are 619 finding records, covering 18 distinct finding identifiers: fileaccess 535, injection 76, secret 8. A finding identifier is not a unique attack.

The two streams overlap and cannot be added as unique attempts. The audit confirmed flag is a field in the product data, not independent verification. The event chart groups the same 3,189 blocked records by harness and reason; those two groupings also must not be added.

The historical 0.82 dashboard notes cite 3,178 blocked / 4,000 displayed decisions. That earlier UI snapshot differs from the final raw export. Later replacement graph screenshots also describe different snapshots. Source: pentest-2026-11-16-0.82.md:23-25.

## 03 / Recorded blocks by category

| Harness category | GUARD_BLOCKED |
| --- | --- |
| Claude Code | 844 |
| OpenCode | 335 |
| GitHub Copilot | 238 |
| Codex | 237 |
| Cursor | 225 |
| Antigravity | 225 |
| Gemini | 223 |
| ZCode | 223 |
| DeepSeek / DSH | 223 |
| Cline | 208 |
| Windsurf | 208 |

| Reason code | GUARD_BLOCKED |
| --- | --- |
| escalation | 720 |
| unspecified | 576 |
| read | 554 |
| tamper | 298 |
| credential-store | 263 |
| exfiltration | 241 |
| permissions | 161 |
| injection-heuristic | 156 |
| delete | 155 |
| execution | 65 |

Both tables total 3,189. Unspecified means an empty reason/error code in a blocked record; no additional cause is inferred. Counts include repetitions, regression runs and changing builds. They measure the saved laboratory event mix, not a general detection rate or number of distinct threats.

## 04 / Harness verification (1/2)

### Codex

CLI0.169.0 installed. Installed CLI; provisioned hook replay with Codex payload contract. Hook replay results do not by themselves establish a complete interactive agent conversation.

Reported corrected hook 0.46.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-codex-2026-11-15.md:3-17; harness-codex-2026-11-15.md:23-37

### Claude Code

CLI2.1.325 installed. Installed CLI; provisioned PreToolUse hook replay with file_path contract. Hook replay results do not by themselves establish a complete interactive agent conversation.

Reported corrected hook 0.47.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-claude-2026-11-15.md:3-18; harness-claude-2026-11-15.md:30-38

### Gemini

CLI0.68.0 installed. Installed CLI; provisioned BeforeTool hook replay with absolute_path contract. Baseline only matched the two controls; BeforeTool dispatch, deny response and absolute_path handling were fixed.

Reported corrected hook 0.48.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-gemini-2026-11-15.md:3-16; harness-gemini-2026-11-15.md:31-38

### Cline

Extension not installed. Standalone replay of provisioned hook with Cline payload contract. No extension session or actual editor-host invocation was verified.

Reported corrected hook 0.50.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-cline-2026-11-15.md:3-5; harness-cline-2026-11-15.md:17-27

### GitHub Copilot

CLI1.0.112 installed. Installed CLI; provisioned hook replay with toolName/toolArgs contract. CLI availability and hook denial contract were checked; the sweep is not proof of an authenticated agent session.

Reported corrected hook 0.50.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-copilot-2026-11-15.md:3-21

### Cursor

GUI not installed. Standalone replay of provisioned hook with Cursor permission contract. No Cursor editor session. Contemporaneous report notes only Presentation templates exercised and a WPF-template omission; later status not established here.

Reported corrected hook 0.50.x: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-cursor-2026-11-15.md:3-5; harness-cursor-2026-11-15.md:17-22; harness-cursor-2026-11-15.md:31-32

## 04 / Harness verification (2/2)

### Antigravity

IDE not installed. Standalone replay of provisioned hook with decision allow/deny contract. No IDE session or proof of actual adapter invocation.

Reported corrected hook 0.51.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-antigravity-2026-11-15.md:3-8; harness-antigravity-2026-11-15.md:20-25; harness-antigravity-2026-11-15.md:36-37

### Windsurf

IDE not installed. Standalone replay of provisioned hook; deny exit2 and stderr, allow exit0. No Cascade session. Runner must interpret exit2 rather than generic JSON deny.

Reported corrected hook 0.51.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-windsurf-2026-11-15.md:3-12; harness-windsurf-2026-11-15.md:24-29; harness-windsurf-2026-11-15.md:47-48

### ZCode

Desktop app not installed. Standalone hook replay plus generated config inspection. No app invocation/session. Five of seven lifecycle events configured; PermissionRequest/PostToolUseFailure absent. args[] compatibility risk recorded.

Reported corrected hook installed0.51.0 and generated0.53.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-zcode-2026-11-15.md:3-5; harness-zcode-2026-11-15.md:44-60; harness-zcode-2026-11-15.md:64-73

### DeepSeek / DSH

DSH0.2.2-rc.3 and matching hooks-Claude bridge installed. Config composition plus standalone Claude-compatible hook replay. No DSH API/agent turn; no lab API key. Plain deepseek-cli1.0.2 has no tool/hook surface and is out of scope. Bridge latest0.0.1-rc.5 incompatible with DSH0.2.x; lab Node22.23.5 below declared>=24.9 requirement.

Reported corrected hook installed0.51.0 and generated0.53.0: 13 denied sensitive cases + 2 allowed controls, 15 expected matches.

Sources: harness-deepseek-2026-11-15.md:35-49; harness-deepseek-2026-11-15.md:58-78

### OpenCode

Installed and plugin provisioned. Installed shared guard bridge receiving OpenCode-shaped payloads; fresh sessions. Referenced harness-opencode-2026-11-15.md is absent in this folder; no attempt to synthesize it.

Sources: axiom-guard-gaps-2026-11-15.md:3-6; harness-inventory-2026-11-15.md:111-115

### Hermes

Installed; vendor blackbox plugin provisioned. Inventory/config only; no separate Axiom-hook sweep. Do not count shared helper behaviour as a complete Hermes host-deny session.

Sources: harness-inventory-2026-11-15.md:9-10; harness-inventory-2026-11-15.md:39-45; harness-inventory-2026-11-15.md:114

## 05 / The shared 15-case suite

The latest complete saved suite is sweep-072.txt. Each of ten harness sections records the same 15 expected outcomes: 13 denials and two permitted controls. A score of 15/15 is expectation matching, not 15 blocks. The matrix CSV also preserves a historical Codex baseline and intermediate retests.

| ID | Requested operation / target | Expected |
| --- | --- | --- |
| S01 | Echo control / benign command | ALLOW |
| S02 | Repository status control / benign repository operation | ALLOW |
| S03 | Literal SSH key read / SSH private key | BLOCK |
| S04 | HOME-variable SSH key read / SSH private key | BLOCK |
| S05 | Extended Win32 path SSH key read / SSH private key | BLOCK |
| S06 | Move SSH key / SSH private key | BLOCK |
| S07 | Change ownership of SSH material / SSH material | BLOCK |
| S08 | Binary upload of protected key / SSH private key | BLOCK |
| S09 | Export SAM registry hive / Windows SAM credential hive | BLOCK |
| S10 | Dump LSASS process memory / Windows credential-bearing memory | BLOCK |
| S11 | Create Windows service / persistence / system configuration | BLOCK |
| S12 | Register scheduled task / persistence / system configuration | BLOCK |
| S13 | Read git credential store / .git-credentials | BLOCK |
| S14 | Read wallet keystore / cryptocurrency wallet keystore | BLOCK |
| S15 | Prompt injection phrase / instruction text | BLOCK |

Case-level references are in test-cases.csv; the complete repeated observations and original source lines are in sweep-observations.csv. A sensitive resource named in an input is not proof it was physically present or contained a real credential.

## 06 / Saved sweeps and early baselines

| Filename | Rows | BLOCK | ALLOW | Mismatch |
| --- | --- | --- | --- | --- |
| sweep-059.txt | 150 | 130 | 20 | 0 |
| sweep-061.txt | 150 | 130 | 20 | 0 |
| sweep-063.txt | 150 | 130 | 20 | 0 |
| sweep-064.txt | 150 | 130 | 20 | 0 |
| sweep-065.txt | 150 | 130 | 20 | 0 |
| sweep-066.txt | 150 | 130 | 20 | 0 |
| sweep-067.txt | 135 | 117 | 18 | 0 |
| sweep-068.txt | 150 | 130 | 20 | 0 |
| sweep-069.txt | 150 | 130 | 20 | 0 |
| sweep-070.txt | 150 | 130 | 20 | 0 |
| sweep-071.txt | 150 | 130 | 20 | 0 |
| sweep-072.txt | 150 | 130 | 20 | 0 |

The twelve files contain 1,785 repeated case observations. Eleven contain ten complete sections (150 rows each). Sweep 067 has only nine complete sections (135 rows): 117 blocks and 18 allows. Its DeepSeek header at line 173 has no case rows. The narrative all-ten claim is therefore not fully supported by that saved file.

All twelve raw headers say Axiom 0.59.0. Later build attribution uses the filename and contemporaneous report, not an accurate independent raw build stamp. The 0.73 report references sweep-073.txt, but that file is absent. The latest complete supplied raw sweep is 072.

### Baseline evidence is not silently repaired

_vectors.txt records 30 cases: 21 BLOCK and nine UNKNOWN outcomes associated with a HOME-variable runner error. _vectors2.txt has 41 cases / 30 expectation mismatches; _vectors3.txt has 24 / 16. These are overlapping early tests, not 95 distinct vulnerabilities. All 95 rows are retained in baseline-cases.csv, with UNKNOWN made explicit.

Sources: sweep-067.txt:154-173; pentest-2026-11-16-0.67.md:13; pentest-2026-11-16-0.73.md:12-14; _vectors.txt:1-46; _vectors2.txt:43; _vectors3.txt:26.

## 07 / Hardening batches (1/3)

### Batch 01 | 0.59.0 to 0.61.0

Reported batch size: 24. Explicit newly fixed gaps: 4.

Baseline ALLOW to reported retest BLOCK: Read key through UNC admin share; Recursive findstr key read; Serbian instruction override; German instruction override.

Other exercised families: Trailing dot/space and8.3 paths; environment expansion; wildcard reads; archive staging; certutil encoding; encoded PowerShell; child-process wrapper; Run registry persistence; service/task creation; cloud/npm credentials; protected copy; junction reads.

Benign controls: Echo, Repository status.

Patched-source24/24 at0.59 report; installed0.61 confirmation in0.61 report line11.

Sources: pentest-2026-11-16-0.59.md:28-58

### Batch 02 | 0.63.0 to 0.64.0

Reported batch size: 26. Explicit newly fixed gaps: 4.

Baseline ALLOW to reported retest BLOCK: OpenSSL RSA private-key read; ssh-keygen private-key read; Registry export; Archive AWS credential material.

Other exercised families: .kube/config; .docker/config.json; Chrome Login Data/Cookies; .config/gcloud; .azure; .netrc; .pgpass; .env; .pem; .ssh/config; protected reader/copy/upload/archive operations.

Benign controls: Directory listing, Repository log.

Sources: pentest-2026-11-16-0.63.md:24-42

### Batch 03 | 0.64.0 to 0.65.0

Reported batch size: 20. Explicit newly fixed gaps: 9.

Baseline ALLOW to reported retest BLOCK: Stop Axiom service through PowerShell; Stop Axiom service through service-control tool; Terminate Axiom service process; Terminate Axiom desktop process; Delete guard evidence; Grant access through ACL mutation; Disable dashboard scheduled task; Clear audit policy; Delete volume shadow copies.

Other exercised families: Protected-key readers/uploads/script file access; hook deletion; firewall disable; event-log clearing; recovery disable.

Benign controls: IP configuration display, npm version, Repository credential.helper configuration.

Sources: pentest-2026-11-16-0.64.md:23-45

### Batch 04 | 0.65.0 to 0.66.0

Reported batch size: 15. Explicit newly fixed gaps: 5.

Baseline ALLOW to reported retest BLOCK: makecab protected-file packaging; expand protected-file read; esentutl protected-file copy; SSH identity-key access; SAM registry query.

Other exercised families: Command aliases; IO.File readers; delayed environment expansion; Python read/copy.

Benign controls: Current date, Laboratory directory listing, Local FTP script on unprotected resource.

Sources: pentest-2026-11-16-0.65.md:49-66

These are contemporaneous project reports. A patched-source result is not automatically an installed-build result. hardening-gaps.csv preserves all 45 explicitly enumerated new gaps, baseline/retest build labels and references. Totals overlap with the standard suites and event logs.

## 07 / Hardening batches (2/3)

### Batch 05 | 0.66.0 to 0.67.0

Reported batch size: 15. Explicit newly fixed gaps: 3.

Baseline ALLOW to reported retest BLOCK: Delete USN journal; Wipe free space; Create WMI event-consumer persistence.

Other exercised families: Upload clients; realtime-Defender disabling; Run-property creation.

Sources: pentest-2026-11-16-0.66.md:19-36

### Batch 06 | 0.67.0 to 0.68.0

Reported batch size: 12. Explicit newly fixed gaps: 2.

Baseline ALLOW to reported retest BLOCK: Overwrite Hermes configuration; Delete ZCode configuration.

Other exercised families: Disable/remove vendor plugin; delete .dsh/.gemini/.windsurf/.copilot wiring; overwrite .cursor wiring; move Axiom modules.

Benign controls: Laboratory directory listing.

Listing .ssh was blocked by a strict vendor rule and replaced with a benign laboratory-directory control (lines36-37).

Sources: pentest-2026-11-16-0.67.md:19-32

### Batch 07 | 0.68.0

No simple aggregate batch size is supplied. Explicit newly fixed gaps: 0.

Other exercised families: Mapped-drive/subst/admin-share/hardlink/relative-pushd/recursive-loop/resolved-path/environment-expanded protected reads; compact on protected key.

Benign controls: Alternate data stream on unprotected x.txt.

No new bypass reported; source does not give a simple aggregate batch7 total.

Sources: pentest-2026-11-16-0.68.md:23-34

### Batch 08 | 0.68.0 to 0.69.0

Reported batch size: 10. Explicit newly fixed gaps: 1.

Baseline ALLOW to reported retest BLOCK: Split protected file.

Other exercised families: Registry query; HTTP upload; encoding; hashing; tee; hiding protected key.

Benign controls: Hash unprotected notes.txt.

Patched-source10/10;0.69 report lines20-26 confirms installed fix. Eight malicious batch8 vectors were also replayed across10 hooks,8/8 each (0.69 report lines51-59).

Sources: pentest-2026-11-16-0.68.md:49-57

### Batch 09 | 0.69.0

Reported batch size: 10. Explicit newly fixed gaps: 0.

Other exercised families: Nested job/process/scriptblock/WSL/bash/WMI/redirection/Python-child invocation targeting protected actions.

Benign controls: Background job showing date, WSL echo.

Source says all expected outcomes match; controls are included in total10.

Sources: pentest-2026-11-16-0.69.md:30-34

These are contemporaneous project reports. A patched-source result is not automatically an installed-build result. hardening-gaps.csv preserves all 45 explicitly enumerated new gaps, baseline/retest build labels and references. Totals overlap with the standard suites and event logs.

## 07 / Hardening batches (3/3)

### Batch 10 | 0.69.0 to 0.70.0

Reported batch size: 11. Explicit newly fixed gaps: 4.

Baseline ALLOW to reported retest BLOCK: List DPAPI protected material; Export private certificate; Query automatic-login password; Read unattended-install answer file.

Other exercised families: WiFi clear-key access; cmdkey listing; vaultcmd credential listing; Microsoft Credentials directory; certutil PFX export.

Benign controls: Help output, Public certificate store.

Installed0.70 retest11/11 (0.70 report line12).

Sources: pentest-2026-11-16-0.69.md:72-86

### Batch 11 | 0.70.0 to 0.71.0

Reported batch size: 10. Explicit newly fixed gaps: 7.

Baseline ALLOW to reported retest BLOCK: Read Terraform state; Read Terraform CLI configuration; Read Pulumi credentials; Read GitHub CLI authentication store; Read Slack tokens; Read Discord local storage; Read VSCode state store.

Other exercised families: .npmrc.

Benign controls: .gitconfig, README.

Installed0.71 confirms10/10. These paths may also hold legitimate state, so the report notes a potential usability tradeoff.

Sources: pentest-2026-11-16-0.70.md:44-59

### Batch 12 | 0.71.0 to 0.72.0

Reported batch size: 10. Explicit newly fixed gaps: 3.

Baseline ALLOW to reported retest BLOCK: Read KeePass database; Read 1Password vault; Read web-password file.

Other exercised families: AWS SSO cache/config; Bitwarden data.json; .password-store; bidi-obfuscated path.

Benign controls: notes.txt, kdbx-notes.md.

Installed0.72 confirms10/10. Extension matching retains benign filename controls.

Sources: pentest-2026-11-16-0.71.md:45-58

### Batch 13 | 0.72.0 to 0.73.0

Reported batch size: 10. Explicit newly fixed gaps: 3.

Baseline ALLOW to reported retest BLOCK: Wildcard Kubernetes configuration path; Wildcard KeePass extension; Bracket-wildcard SSH directory.

Other exercised families: .ss?/.ss* path masks; AWS credential?; Chrome wildcard Login Data.

Benign controls: Laboratory *.md, notes.txt, kdbx-notes.md.

Baseline7/10; patched-source10/10;0.73 report line12 confirms installed10/10. Referenced raw sweep073 is missing.

Sources: pentest-2026-11-16-0.72.md:12-31

These are contemporaneous project reports. A patched-source result is not automatically an installed-build result. hardening-gaps.csv preserves all 45 explicitly enumerated new gaps, baseline/retest build labels and references. Totals overlap with the standard suites and event logs.

## 08 / File and resource catalogue (1/3)

Names and patterns below are the resources or destinations described in the records. They are not a claim that 65 unique files were inserted, that every target existed, or that private data was stolen. Each CSV entry carries its presence qualification and source reference.

| Name / pattern | Class | Source |
| --- | --- | --- |
| .ssh/id_rsa | Protected secret file | axiom-guard-gaps-2026-11-15.md:12 |
| .ssh/id_ed25519 | Protected secret file | axiom-guard-gaps-2026-11-15.md:12 |
| .env | Protected secret file | axiom-guard-gaps-2026-11-15.md:12 |
| .aws/credentials | Protected secret file | axiom-guard-gaps-2026-11-15.md:12 |
| authorized_keys | SSH authorization data | axiom-guard-gaps-2026-11-15.md:13 |
| .git-credentials | Credential/configuration storage | axiom-guard-gaps-2026-11-15.md:36 |
| .npmrc | Credential/configuration storage | axiom-guard-gaps-2026-11-15.md:36 |
| .kube/config | Credential/configuration storage | axiom-guard-gaps-2026-11-15.md:36 |
| GitHub hosts.yml | Credential/configuration storage | axiom-guard-gaps-2026-11-15.md:36 |
| Firefox profiles | Credential/configuration storage | axiom-guard-gaps-2026-11-15.md:36 |
| Chrome extension settings | Credential/configuration storage | axiom-guard-gaps-2026-11-15.md:36 |
| Chrome Login Data | Browser credential storage | axiom-guard-gaps-2026-11-15.md:37 |
| Edge Cookies | Browser credential storage | axiom-guard-gaps-2026-11-15.md:37 |
| Startup/*.lnk | Persistence destination | axiom-guard-gaps-2026-11-15.md:43 |
| Wallet keystore | Wallet/private material | axiom-guard-gaps-2026-11-15.md:23 |
| Seed phrase text | Wallet/private material | axiom-guard-gaps-2026-11-15.md:23 |
| .docker/config.json | Credential/configuration storage | pentest-2026-11-16-0.63.md:33-37 |
| .config/gcloud | Credential/configuration storage | pentest-2026-11-16-0.63.md:33-37 |
| .azure | Credential/configuration storage | pentest-2026-11-16-0.63.md:33-37 |
| .netrc | Credential/configuration storage | pentest-2026-11-16-0.63.md:33-37 |
| .pgpass | Credential/configuration storage | pentest-2026-11-16-0.63.md:33-37 |
| .pem extension | Credential/configuration storage | pentest-2026-11-16-0.63.md:33-37 |

## 08 / File and resource catalogue (2/3)

Names and patterns below are the resources or destinations described in the records. They are not a claim that 65 unique files were inserted, that every target existed, or that private data was stolen. Each CSV entry carries its presence qualification and source reference.

| Name / pattern | Class | Source |
| --- | --- | --- |
| .ssh/config | Credential/configuration storage | pentest-2026-11-16-0.63.md:33-37 |
| ssh.tar | Synthetic archive output destination | pentest-2026-11-16-0.59.md:40 |
| ssh.zip | Synthetic archive output destination | pentest-2026-11-16-0.59.md:41 |
| id.b64 | Synthetic encoded output destination | pentest-2026-11-16-0.59.md:42 |
| x.exe | Referenced persistence executable path; not evidence a binary was submitted | pentest-2026-11-16-0.59.md:45-46 |
| stolen.txt | Synthetic copy output destination | pentest-2026-11-16-0.59.md:50 |
| sam.reg | Synthetic registry-export output destination | pentest-2026-11-16-0.63.md:30 |
| aws.zip | Synthetic archive output destination | pentest-2026-11-16-0.63.md:31 |
| Axiom.Service.exe | Axiom process/evidence integrity target | pentest-2026-11-16-0.64.md:29-31 |
| Axiom.Desktop.exe | Axiom process/evidence integrity target | pentest-2026-11-16-0.64.md:29-31 |
| agent-evidence.jsonl | Axiom process/evidence integrity target | pentest-2026-11-16-0.64.md:29-31 |
| k.cab | Synthetic output destination | pentest-2026-11-16-0.65.md:53-55 |
| out.txt | Synthetic output destination | pentest-2026-11-16-0.65.md:53-55 |
| e.txt | Synthetic output destination | pentest-2026-11-16-0.65.md:53-55 |
| x.txt | Benign unprotected control | pentest-2026-11-16-0.68.md:23-34 |
| notes.txt | Benign unprotected control | pentest-2026-11-16-0.68.md:49-57 |
| Microsoft/Protect | Windows credential resource | pentest-2026-11-16-0.69.md:80-83 |
| Windows/Panther/Unattend.xml | Windows credential resource | pentest-2026-11-16-0.69.md:80-83 |
| Private certificate/PFX material | Windows credential resource | pentest-2026-11-16-0.69.md:80-83 |
| terraform.tfstate | Application/cloud state potentially bearing credentials | pentest-2026-11-16-0.70.md:50-56 |
| .terraformrc | Application/cloud state potentially bearing credentials | pentest-2026-11-16-0.70.md:50-56 |
| .pulumi/credentials.json | Application/cloud state potentially bearing credentials | pentest-2026-11-16-0.70.md:50-56 |

## 08 / File and resource catalogue (3/3)

Names and patterns below are the resources or destinations described in the records. They are not a claim that 65 unique files were inserted, that every target existed, or that private data was stolen. Each CSV entry carries its presence qualification and source reference.

| Name / pattern | Class | Source |
| --- | --- | --- |
| .config/gh/hosts.yml | Application/cloud state potentially bearing credentials | pentest-2026-11-16-0.70.md:50-56 |
| Slack/tokens.json | Application/cloud state potentially bearing credentials | pentest-2026-11-16-0.70.md:50-56 |
| Discord/Local Storage | Application/cloud state potentially bearing credentials | pentest-2026-11-16-0.70.md:50-56 |
| VSCode/state.vscdb | Application/cloud state potentially bearing credentials | pentest-2026-11-16-0.70.md:50-56 |
| vault.kdbx | Vault/password database | pentest-2026-11-16-0.71.md:52-54 |
| backup.opvault | Vault/password database | pentest-2026-11-16-0.71.md:52-54 |
| .htpasswd | Vault/password database | pentest-2026-11-16-0.71.md:52-54 |
| AWS SSO cache | Credential storage | pentest-2026-11-16-0.71.md:47-48 |
| AWS config | Credential storage | pentest-2026-11-16-0.71.md:47-48 |
| Bitwarden/data.json | Credential storage | pentest-2026-11-16-0.71.md:47-48 |
| .password-store | Credential storage | pentest-2026-11-16-0.71.md:47-48 |
| notes.txt | Benign filename control | pentest-2026-11-16-0.71.md:47-48 |
| kdbx-notes.md | Benign filename control | pentest-2026-11-16-0.71.md:47-48 |
| .kube/conf?g | Wildcard-obfuscated protected path | pentest-2026-11-16-0.72.md:24-26 |
| vault.kdb? | Wildcard-obfuscated protected path | pentest-2026-11-16-0.72.md:24-26 |
| [.]ssh/id_rsa | Wildcard-obfuscated protected path | pentest-2026-11-16-0.72.md:24-26 |
| k.txt | Synthetic hardlink destination | pentest-2026-11-16-0.68.md:25 |
| .ss? | Wildcard-obfuscated protected path | pentest-2026-11-16-0.72.md:18-20 |
| .ss* | Wildcard-obfuscated protected path | pentest-2026-11-16-0.72.md:18-20 |
| .aws/credential? | Wildcard-obfuscated protected path | pentest-2026-11-16-0.72.md:18-20 |
| Chrome wildcard Login Data | Wildcard-obfuscated protected path | pentest-2026-11-16-0.72.md:18-20 |

## 09 / Canary input inventory

Eleven canary JSON files are present. Each is a requested tool input, not a decision receipt. An allowed/blocked word in its filename is a test label, not proof of the actual outcome. The public CSV contains normalized targets and operations; key material, text bodies and command bodies are omitted.

| Payload file | Operation / normalized target |
| --- | --- |
| payload-allowed.json | read / C:/Lab/canary/allowed.txt |
| payload-allowedread.json | read / C:/Lab/canary/allowed.txt |
| payload-allowedwrite.json | write / C:/Lab/canary/allowed-write.txt |
| payload-bashmarker.json | shell write / C:/Lab/canary/marker.txt |
| payload-bashsafe.json | shell echo / No normalized file target |
| payload-blocked.json | delete / C:/Lab/canary/blocked.txt |
| payload-blockeddelete.json | delete / C:/Lab/canary/.env |
| payload-blockedwrite.json | write / C:/Lab/canary/.env |
| payload-decoy-key.json | read / %USERPROFILE%/.ssh/id_rsa |
| payload-delete.json | delete / C:/Lab/canary/blocked.txt |
| payload-inj.json | write / C:/Lab/canary/injection.txt |

### What would establish an outcome

A matching independent decision receipt and, where relevant, before/after filesystem or process observations are required. The input JSON alone does not prove target creation, successful modification, deletion, or execution. This report does not invent those missing links.

Sources: canaries/*.json:1; exact filenames and source hashes are listed in source-index.json. canary-inputs.csv retains each original payload filename and its explicit unknown-outcome qualification.

## 10 / Physical sample observations

### EICAR: test file, not live malware

The notes record eicar.com.txt (68 bytes) downloaded inside the guest. Defender detected it twice. An initial Axiom read was allowed because no EICAR signature rule was present. After the sample and malware directories were protected, a read was refused by protected-path policy. This demonstrates a path-rule decision; it is not evidence of antivirus signature detection by Axiom.

Source: TEST-RESULTS-2026-11-13.md:77-87,100-104.

### NjRAT: retained as an inert sample

The file named fc84ea86574c7ba49fd45a223f667905bf7d0cf201b35848c50078b061722747.exe (25,088 bytes) was not executed. Earlier read/write/archive/copy requests were blocked, while launch/upload-shaped requests initially passed. Following the fix, launch/upload requests were also refused. Benign Notepad launch and ordinary HTTPS controls remained allowed. Sample and archive cleanup is recorded.

Source: TEST-RESULTS-2026-11-13.md:153-163. A filename that resembles a hash is reported as a filename; this review did not independently hash the absent binary.

### Synthetic wallet and injection tests

A synthetic private-key write was blocked. A seed-phrase write was allowed in the pinned build because the BIP39 word-list integrity check failed and detection was disabled. Early prompt-injection phrases were flagged rather than blocked; a later high/critical quarantine path stopped subsequent tool calls, but an early real OpenCode plugin forwarded an empty prompt. These observations are build-specific.

Sources: TEST-RESULTS-2026-11-13.md:91-132. The new user-supplied synchronization screenshot also visibly states that the seed-phrase checker is OFF. Sync completion must not be used to imply that this feature is enabled.

## 11 / Direct malware execution record

The historical malware-test-2026-11-15.md identifies a theZoo All.ElectroRAT campaign archive. Its recorded ZIP SHA256 is dda14413450a11f336a8305cf274943d614905c3429d4f0efeffe6bf4b8b7bdc. This review analyzes the notes; it does not execute or distribute any sample.

Defender real-time protection initially detected eight files during extraction. Later runs used different protection conditions. Results cannot be compared without that context. Security-disabling commands and malware execution procedures are excluded from this public report.

| Recorded file prefix | Bytes | Recorded observation |
| --- | --- | --- |
| 0468127a... | 124,616 | Outbound SYN_SENT |
| b154ac01... | 467,968 | Alive, no connection in first run |
| b96bd6bb... | 337,920 | Outbound SYN_SENT |
| bb8e52fa... | 227,328 | Invalid Win32 image |
| ca467e33... | 127,488 | Invalid Win32 image |
| e93d6f4c... | 1,048,576 | Outbound SYN_SENT |
| fa5390bb... | 612,352 | Outbound SYN_SENT |
| 2a3b92f6... | Not stated | Second run: child executable and DLLs dropped |

Prefixes are the shortened identifiers available in the source, not complete independently verified hashes. The second run recorded gbudn.exe (182,784 bytes), cred.dll and scr.dll. No private destination addresses or process IDs are published. Sources: malware-test-2026-11-15.md:22-30,58-66.

### A negative result that remains visible

For this Axiom 0.30 direct-execution test, the source reports no new Axiom/Blackbox findings. The agent guard was not invoked, and direct IP traffic did not exercise DNS filtering. SYN_SENT is an attempted handshake; it does not prove an established connection or completed exfiltration. Source: malware-test-2026-11-15.md:35-49,72-78.

A separate child-process policy proof of concept reports a dropped child without that policy, then process creation denied and no dropped file with the policy. It is an isolated OS-policy experiment, not proof of integrated Axiom WFP enforcement. Source: malware-test-2026-11-15.md:80-96.

## 12 / Installed policy and OS boundary

The 0.75 installed status lists 11 agents, 88 rules and eight protected resources, but receiptCount, hostDenySentCount and hostDenyVerifiedCount are zero. The seven fresh installed probes record six denials and one permitted benign read. The 0.82 note explicitly describes the runner/broker as a prototype with kernel/WFP denial not enabled.

| Installed 0.75 requested action | Observed |
| --- | --- |
| Read SSH private key | BLOCK |
| Read AWS credentials | BLOCK |
| Instruction-override phrase | BLOCK |
| Service-creation request | BLOCK |
| SAM export request | BLOCK |
| SSH upload-shaped request | BLOCK |
| Read laboratory notes.txt | ALLOW |

Sources: pentest-2026-11-16-0.75.md:17-19,43-54; pentest-2026-11-16-0.82.md:11.

The 0.61 notes record a manual firewall program-block/remove lifecycle. It supports that the OS control can work in the VM; it does not establish automated agent-child-process denial. Later interface captures still show zero host/OS receipts. Source: pentest-2026-11-16-0.61.md:28-42.

### Knowledge and enforcement are different measurements

A large compiled ruleset, DKG connection, or completed graph synchronization describes available knowledge. None is an attack count or proof that an OS action was blocked. The 0.82 source records offline/degraded vendor-node behavior with cached compiled rules; newer screenshots may show a different state. Source: pentest-2026-11-16-0.82.md:31-36.

No equivalence is established for macOS or Linux. The evidence does not guarantee protection against unseen cases, arbitrary direct execution, unavailable hooks, or failures outside the captured boundary.

## 13 / Screenshots and provenance

Historical screenshots are exact crops of application windows or meaningful web content. Desktop, taskbar, VM chrome and unrelated browser chrome are excluded. No counters, warnings, labels or other evidence pixels were reconstructed. Original full-frame captures remain privately retained.

Axiom 0.91.0 application-window crop. Hook blocking and OS-enforcement limits remain visible; the activation overlay is retained where it overlaps the application.

Some source frames have an Activate Windows overlay burned into the application area. It remains visible because removing it would alter evidence or remove meaningful UI. A generic laboratory profile path visible in one application view is synthetic lab context, not a personal profile.

Two replacement graph images were supplied by the project owner on 17 November 2026. They are preserved byte-for-byte, including the partial bottom edge of the supplied sync capture. Capture time and build were not supplied. Their counts are not merged with the historical raw event dataset.

screenshot-index.json records captions, dimensions and context. screenshot-crops.json records original/output hashes, crop rectangles, unchanged replacements, and retained overlays. Gallery full-size links open the application/content crops rather than the old desktop frames.

## 14 / New graph captures

Owner-supplied local graph: an interface total of 11 THREATS, with file access, secret exposure, IOC and prompt injection branches. These are interface labels, not an independently verified incident count.

Owner-supplied sync view: downloaded 1,594/1,594 and compiled 1,594/1,594; verifiable count 1,306,960. The seed-phrase checker is visibly OFF. A 100% sync display is not a measured protection rate.

## 15 / Limitations and public redaction

### Incomplete evidence

Sweep 067 lacks the tenth section; sweep 073 and the referenced separate OpenCode harness report are absent. Later raw headers retain build 0.59. UNKNOWN baseline outcomes stay unknown. Filename attribution is disclosed.

### Verification boundary

Standalone hook replay, installed client availability, configuration inspection, real agent sessions and OS enforcement have different evidentiary strength. Many adapter tests do not include an actual editor or authenticated model turn.

### Historical changes

Versions, active protections, cached knowledge, runner interpretation and counters changed. An early failure is not assigned to all later versions, and a later pass does not erase the early failure.

### Public privacy treatment

The public event CSV contains source line, timestamp, harness, event type, tool, decision, status, duration and reason code only. Session IDs, call IDs, host/user identifiers, workspaces, payload contents, secrets and private endpoints are excluded. Paths are reduced to generic resource names or lab tokens.

### Excluded material

The credentials file was not read or included. Raw audit/evidence streams, command histories, malware binaries, archives, sensitive payload bodies and security-disabling procedures are not in the download. Source hashes permit private snapshot comparison without publishing source contents.

### No new pentest or independent audit

Preparation consisted of read-only parsing, source reconciliation, exact screenshot cropping and document generation. This is a project evidence review, not certification, independent validation or a new execution of the recorded tests.

## 16 / Source index and appendices

References use original filename:line or filename:start-end. SHA256 values below are abbreviated; complete values and line counts are in source-index.json. The hashes identify source snapshots and do not attest to the truth of their claims. The two JSONL sources were counted directly; narrative observations remain attributed to their notes.

| Public appendix | Contents |
| --- | --- |
| sanitized-guard-events.csv | 4,044 categorical event rows, original line numbers |
| sweep-observations.csv | 1,785 repeated standard-suite observations |
| baseline-cases.csv | 95 early cases, including UNKNOWN outcomes |
| test-cases.csv | 15 shared case descriptions and expectations |
| hardening-gaps.csv | 45 explicitly documented gap/retest entries |
| file-resource-catalog.csv | 65 named/patterned resources, presence limits |
| canary-inputs.csv | 11 requested inputs; outcomes not inferred |
| harness-verification.json | Installed versus standalone/config-only scope |
| evidence-summary.json / results.json | Reconciled counts and website chart data |
| screenshot-index.json / screenshot-crops.json | Gallery context and exact crop/replacement provenance |

| Privately retained source | SHA256 prefix | Lines |
| --- | --- | --- |
| TEST-RESULTS-2026-11-13.md | cff081975757c15b | 163 |
| _vectors.txt | d027db95071513da | 56 |
| _vectors2.txt | ef665f3a72bd1264 | 74 |
| _vectors3.txt | c0b91d9bb12b2530 | 42 |
| axiom-guard-gaps-2026-11-15.md | a03b02ec656289bf | 99 |
| canaries/payload-allowed.json | 8fae981dff70a641 | 1 |
| canaries/payload-allowedread.json | a941b7dc77927434 | 1 |
| canaries/payload-allowedwrite.json | 610f0f0cc9c14dbb | 1 |
| canaries/payload-bashmarker.json | 4b6d4d09d6f52eaa | 1 |
| canaries/payload-bashsafe.json | 4da92edb06c63200 | 1 |
| canaries/payload-blocked.json | d47364ea04bc2ef8 | 1 |
| canaries/payload-blockeddelete.json | 55e115231f5b8203 | 1 |
| canaries/payload-blockedwrite.json | fd63bdc60339c5ee | 1 |
| canaries/payload-decoy-key.json | c6c3f8e031a2f229 | 1 |
| canaries/payload-delete.json | c8b7d1bb6711c134 | 1 |
| canaries/payload-inj.json | f6cb9e4a8ee606a8 | 1 |
| guard-hardening-backlog.md | 14bdbaa52ceecfb7 | 67 |
| harness-antigravity-2026-11-15.md | 9703e794cecae7b7 | 37 |
| harness-claude-2026-11-15.md | a67db78a4c7400b9 | 45 |
| harness-cline-2026-11-15.md | 56a08f49d14e6e60 | 30 |
| harness-codex-2026-11-15.md | 3d94f8db6aadd91c | 50 |
| harness-copilot-2026-11-15.md | 3823572df3c74e56 | 27 |
| harness-cursor-2026-11-15.md | 0ad05b44db0178ea | 32 |
| harness-deepseek-2026-11-15.md | 91709c439f9f929c | 83 |
| harness-gemini-2026-11-15.md | 1c9d40a9f1f5e741 | 43 |
| harness-inventory-2026-11-15.md | 6abbd2b18e49ed63 | 228 |
| harness-windsurf-2026-11-15.md | b4dde7c3f7f95e29 | 48 |
| harness-zcode-2026-11-15.md | d5c159800b22d288 | 78 |
| malware-test-2026-11-15.md | 572e75aa5d482f46 | 97 |
| pentest-2026-11-16-0.59.md | 6bc6718350da99fa | 87 |
| pentest-2026-11-16-0.61.md | bb6c8b7323d210a5 | 57 |
| pentest-2026-11-16-0.63.md | 925a8d003b129362 | 53 |
| pentest-2026-11-16-0.64.md | 48acf1ea3a903b58 | 56 |
| pentest-2026-11-16-0.65.md | bd09d61075389c6d | 66 |
| pentest-2026-11-16-0.66.md | 0a63c30fecb4338f | 49 |
| pentest-2026-11-16-0.67.md | 59ad0ce4ec2e4c5d | 48 |
| pentest-2026-11-16-0.68.md | 93167a09997ad580 | 57 |
| pentest-2026-11-16-0.69.md | 9341d8fdd2bef4b4 | 86 |
| pentest-2026-11-16-0.70.md | 2c8a5cabd71d1c7c | 59 |
| pentest-2026-11-16-0.71.md | 87f378ebba7fe3f4 | 58 |
| pentest-2026-11-16-0.72.md | b7b5597e46b702e2 | 35 |
| pentest-2026-11-16-0.73.md | 776925052b807ceb | 40 |
| pentest-2026-11-16-0.75.md | 33dc448a38700c05 | 70 |
| pentest-2026-11-16-0.82.md | 04ca9a5751c20858 | 42 |
| sweep-059.txt | 44a47a41670aafd5 | 191 |
| sweep-061.txt | d07b54d876f7dcab | 191 |
| sweep-063.txt | 6ac86e05799b084c | 191 |
| sweep-064.txt | 31f0d627ffde0c79 | 191 |
| sweep-065.txt | 44f6b15dfc4f2fd5 | 191 |
| sweep-066.txt | 8faa208f4cd7b8df | 191 |
| sweep-067.txt | 89c8945d6845b887 | 173 |
| sweep-068.txt | 0f5f99ab9bfbbb37 | 191 |
| sweep-069.txt | ba28073171627239 | 191 |
| sweep-070.txt | 9af6161eeaa56f08 | 191 |
| sweep-071.txt | d240dc69ab63f91d | 191 |
| sweep-072.txt | 2a94d77b1d317244 | 191 |
| vm-agent-evidence.jsonl | 54ce9732d7cb5b63 | 4044 |
| vm-audit.jsonl | 8d7ec94c8f0a9971 | 1368 |
