[integrity-audit] Integrity Filtering Audit — clean bill of health (2026-08-20/21)

[github-actions]

Integrity Filtering Audit — github/gh-aw-mcpg

Audit period: Last 24 hours (2026-08-20T06:02 – 2026-08-21T06:02 UTC)
Runs analyzed: 28 completed runs in github/gh-aw (main branch + several PR/copilot branches)
Runs with artifacts inspected: 4 (deep-dive on failures) + spot-checks on recent successes

Findings Summary

Severity	Count	Description
🔴 Critical	0	No data leaks, guard bypasses, or labeling failures found
🟡 Warning	0	No over-filtering, unscoped tags, WASM traps, or direct API bypass attempts found
🟢 Info	5	Infra-related failures unrelated to DIFC; normal gateway/DIFC behavior confirmed

Critical Findings

None found.

Warnings

None found. No firewall access-log entries pointed at api.github.com, github.com, chatgpt.com, openai.com, or any other non-allowlisted domain in the four artifact bundles inspected. All GitHub tool traffic in mcp-gateway.log / rpc-messages.jsonl went through the github MCP backend server (command=docker, routed via awmg-mcpg:8080/mcp/github) as expected, with no unknown REST endpoint, guard not initialized, or wasm error: entries.

Informational

• Daily VulnHunter Scan (run 32451994270) — failed. Root cause: Cloud Hypervisor guest connectivity probe failed with exit code 4 in the sandbox networking layer, before the MCP Gateway ever received meaningful traffic. agent_output.json was empty ({"items":[],"errors":[]}), rpc-messages.jsonl shows only the two normal tools/list exchanges (safeoutputs + github) needed for tool registration, zero DIFC filter events. This is a runner/sandbox networking issue, not a DIFC problem.
• Daily Container Image Security Scan (run 32451259577) — same Cloud Hypervisor guest-connectivity failure signature as VulnHunter above. Tool registration completed (36 tools), no DIFC events logged, no guard/gateway errors, no firewall bypass attempts.
• GitHub Remote MCP Authentication Test (run 32451067175) — same Cloud Hypervisor guest-connectivity failure. Gateway started, registered github and safeoutputs MCP servers, negotiated tools/list and prompts/list successfully before the guest network dropped; no anomalies in the RPC trace captured.
• Daily Sub-Agent Model Resolution Audit (run 32451181394) — failed with Error [0]: context deadline exceeded on an agenticworkflows MCP logs tool call that ran 60s (client-side timeout), unrelated to DIFC/guard enforcement. No firewall blocks, no guard errors.
• Activation rate-limit failures: none observed among the 28 runs sampled in the last 24h; all completed runs reached the agent/gateway stage.

Across all four inspected artifact bundles: zero difc_integrity/difc_secrecy tag entries were logged in rpc-messages.jsonl for these particular runs (github/safeoutputs tools/list exchanges only — no data-fetching tool calls occurred before the runs failed), so there was no filterable GitHub API response data to evaluate for scope violations in this window. This is expected given the runs failed early (either during sandbox network bring-up or during a slow subprocess call), not due to DIFC.

Runs Analyzed

Run	Workflow	Branch	Agent Invoked	DIFC Events	Firewall Blocks	Status
32451994270	Daily VulnHunter Scan	main	✅	0	0/3	⚠️ (infra failure)
32451259577	Daily Container Image Security Scan	main	✅	0	0/3	⚠️ (infra failure)
32451181394	Daily Sub-Agent Model Resolution Audit	main	✅	n/a	0	⚠️ (tool timeout)
32451067175	GitHub Remote MCP Authentication Test	main	✅	n/a	0	⚠️ (infra failure)
32452150508	Daily PR Code Quality Review	main	✅	—	—	✅ success
32451933759	Daily Safe Outputs Conformance Checker	main	✅	—	—	✅ success
32451842844	Issue Monster	main	✅	—	—	✅ success
32451362435	PR Sous Chef	main	✅	—	—	✅ success
32451206866	Avenger	main	✅	—	—	✅ success
32451192432	Multi-Device Docs Tester	main	✅	—	—	✅ success
32450908607	Schema Consistency Checker	main	✅	—	—	✅ success
32450955372	CLI Version Checker	main	—	—	—	❌ failure (not deep-dived; no MCP Gateway artifact pattern observed)

(Remaining action_required runs are workflow-approval gated PR/dispatch runs from copilot/* branches that never executed the agent job, so they carry no MCP/DIFC data to audit.)

Recommendations

1. No DIFC/guard code changes recommended this cycle — enforcement pipeline is behaving as expected (clean tool registration, no unscoped tags, no bypass attempts) in all inspected runs.
2. The recurring Cloud Hypervisor guest-connectivity probe failed with exit code 4 failure (seen in 3 of 4 deep-dived runs: VulnHunter, Container Image Security Scan, GitHub Remote MCP Authentication Test) looks like a shared-runner networking/sandbox provisioning issue rather than an application bug. Recommend filing/tracking this separately with the runner/sandbox infrastructure team, since it is blocking multiple daily scheduled workflows from producing any output.
3. Investigate the agenticworkflows logs MCP tool's 60s timeout behavior (context deadline exceeded) in the Sub-Agent Model Resolution Audit run — consider raising the client timeout or adding pagination/streaming for large log downloads so slow gh run download calls don't hard-fail the tool call.
4. Continue the current cadence: no tools.github misconfiguration or prompt-hardening changes needed at this time since no workflow exhibited direct api.github.com/github.com/external-AI-service bypass attempts.

Generated by Integrity Filtering Audit · auto · 112.5 AIC · ⊞ 13.4K · ◷

• expires on Aug 28, 2026, 6:06 AM UTC