[security-observability] Daily Security Observability Report — 2026-09-02

[github-actions[bot]]

Executive Summary

This report combines two complementary security signals across the last 7 days for the github/gh-aw repository. Note on data window: the audit tooling returns the 100 most recent firewall-enabled runs and 12 most recent DIFC-filtered runs available; in practice these spanned a much shorter recent window (roughly a few hours to ~2 days) rather than the full 7 days, since the repository runs a very high volume of agentic workflows. The analysis below reflects that available data faithfully.

Firewall telemetry across 92 analyzed workflow runs shows a low overall block rate (1.56%) with 90 blocked requests out of 5,752 total. Blocking activity concentrates on a small set of domains — most notably ab.chatgpt.com:443 (40 blocks, likely a Copilot/OpenAI telemetry or analytics endpoint not currently allowlisted) and github.com:443 (30 blocks, unusual since GitHub API access is normally expected — worth investigating whether these are non-API paths or rate-limited calls). No large-scale exfiltration or clearly malicious domains were observed.

On the DIFC side, 410 integrity-filtered tool-call events were recorded across 8 distinct workflows, overwhelmingly on list_issues (260) and search_issues (133) calls against the github MCP server. All filtering was attributed to the integrity category (no secrecy-tag filtering observed), driven by none:public (410) and unapproved:public (287) integrity tags — indicating the DIFC policy is consistently blocking access to public GitHub issues/PRs that haven't been explicitly approved for agent consumption. Workflow Normalizer (179 events) and the Smoke Agent: public/approved test workflow (130 events) account for the majority of filtered activity, with the latter being expected smoke-test behavior validating the DIFC boundary itself.

---

🔥 Firewall Analysis

Key Firewall Metrics

Metric	Value
Workflows analyzed (firewall-enabled)	92 (of 100 collected; 8 runs had no firewall_analysis data)
Total network requests monitored	5,752
✅ Allowed requests	5,662
🚫 Blocked requests	90
Block rate	1.56%
Total unique blocked domains	7

📈 Firewall Request Trends

The available data window spans roughly 4 hours (11:00–15:00 UTC on 2026-09-02), reflecting the volume of the 100 most recent firewall-enabled runs rather than a full calendar week. Blocked request volume tracks proportionally with allowed volume (peaking together around 13:00–14:00 UTC), suggesting blocks are a steady background rate tied to overall workflow throughput rather than a sudden spike or attack pattern.

Top Blocked Domains

ab.chatgpt.com:443 and github.com:443 dominate blocked traffic. The former appears repeatedly across many unrelated workflows (AI Moderator, Issue Monster, PureLock, Workflow Normalizer, etc.), suggesting it's likely a common OpenAI/ChatGPT client-side analytics or update-check endpoint triggered by an underlying tool/library rather than a targeted security concern. github.com:443 blocks across PR-focused workflows (PR Code Quality Reviewer, PR Sous Chef, Test Quality Sentinel) warrant a closer look — these may be legitimate GitHub API/web calls being blocked due to path or method restrictions in the firewall policy.

Most Frequently Blocked Domains

Domain	Times Blocked	Workflows	Category
ab.chatgpt.com:443	40	AI Moderator, Daily Credit Limit Test, Daily Go Test Parallelizer, Issue Monster, Issue Triage Agent, Ponytail Reviewer, PureLock, Workflow Normalizer	Telemetry/Analytics
github.com:443	30	PR Code Quality Reviewer, PR Sous Chef, Test Quality Sentinel	GitHub API/Web
(unknown)	9	Architecture Guardian, CLI Consistency Checker, Daily Copilot PR Merged Report, Daily Credit Limit Test, Daily Go Test Parallelizer, Issue Triage Agent, PureLock, Workflow Normalizer	Unresolved
api.anthropic.com:443	4	Claude Code User Documentation Review, Deep Report	AI Provider API
registry.npmjs.org:443	2	Impeccable Skills Reviewer, Matt Pocock Skills Reviewer	Package Registry
storage.googleapis.com:443	2	Delight	Cloud Storage
proxy.golang.org:443	2	Daily Formal Spec Verifier	Package Proxy
and	1	Deep Report	Malformed/Parsing Artifact

Policy Rule Attribution

📋 Policy: 11 rules, SSL Bump disabled, DLP disabled

Rule-level hit/deny attribution (policy_analysis.rule_hits) was not populated in the audit data for these runs — the policy summary confirms an 11-rule egress policy is active, but per-rule counters were empty across all 92 audited runs. This is worth flagging to the audit tooling maintainers as a potential gap in policy_analysis reporting.

View Detailed Request Patterns by Workflow

Blocking is concentrated in a long tail of individual runs rather than any single workflow dominating overall traffic. Workflows with multiple firewall-enabled runs in this window (e.g., PR Sous Chef, Issue Monster, PR Code Quality Reviewer) show occasional single-digit blocks per run, consistent with incidental blocked calls (e.g., telemetry beacons) rather than systematic policy violations. No workflow showed blocked requests exceeding ~10% of its total request volume.

View Complete Blocked Domains List

• and (likely a log-parsing artifact, not a real domain)
• ab.chatgpt.com:443
• api.anthropic.com:443
• github.com:443
• proxy.golang.org:443
• registry.npmjs.org:443
• storage.googleapis.com:443
• (unknown) — domain could not be resolved from the log entry

🔒 Firewall Security Recommendations

1. Allowlist ab.chatgpt.com:443 if it is confirmed to be a benign OpenAI/ChatGPT client telemetry endpoint used by the Copilot/AI engine runtime — this would eliminate the largest source of blocked-request noise (40 of 90 blocks).
2. Investigate github.com:443 blocks in PR-review workflows (PR Code Quality Reviewer, PR Sous Chef, Test Quality Sentinel) — confirm whether these are legitimate github.com web-page fetches (as opposed to api.github.com) that should be added to the network allowlist, or whether they represent an unexpected/undesired access pattern.
3. Fix the (unknown)/and domain-logging artifacts — 9 blocks resolve to (unknown) and 1 to the literal string and, indicating a log-parsing or domain-extraction bug in the firewall log pipeline that should be reported upstream.
4. No action needed for api.anthropic.com:443, registry.npmjs.org:443, storage.googleapis.com:443, proxy.golang.org:443 — low-frequency, workflow-appropriate (AI provider, npm, GCS, Go proxy) blocks that are expected under a default-deny egress policy; add to allowlists only if a specific workflow requires ongoing access.
5. Request policy_analysis.rule_hits population from the audit tooling team — rule-level attribution was empty across all 92 runs, limiting the ability to tie specific blocks back to specific firewall policy rules.

---

🔒 DIFC Integrity Analysis

Key DIFC Metrics

Metric	Value
Total filtered events	410
Unique tools filtered	6
Unique workflows affected	8
Most common filter reason	integrity (100%)
Busiest day	2026-09-01 (354 events)

📈 DIFC Events Over Time

Filtering activity is heavily front-loaded on 2026-09-01 (354 of 410 events, 86%), dropping to 56 events on 2026-09-02. The 2026-09-01 spike aligns with the Smoke Agent test suite and Workflow Normalizer runs, both of which intentionally exercise the DIFC boundary (smoke tests) or process large volumes of public issues (Workflow Normalizer), rather than indicating a security incident.

🔧 Top Filtered Tools

list_issues (260) and search_issues (133) together account for 96% of all filtered calls. Both tools return bulk collections of GitHub issues, and the DIFC integrity boundary is correctly blocking access to unapproved public issue content within those result sets — this is the expected behavior for an integrity-below-threshold read pattern, not a misconfiguration.

🏷️ Filter Reasons and Tags

100% of filtered events were categorized as integrity filtering; no secrecy-tag filtering occurred in this window. Within integrity tags, none:public (410, i.e. every filtered event) and unapproved:public (287) dominate, with only 7 events tagged approved:public (likely borderline cases where partial approval existed but another integrity constraint still triggered filtering).

📋 Per-Workflow DIFC Breakdown

Workflow	Filtered Events
Workflow Normalizer	179
Smoke Agent: public/approved	130
Smoke Temporary ID	47
Issue Triage Agent	23
Smoke Agent: all/merged	13
GitHub Remote MCP Authentication Test	10
Smoke Agent: scoped/approved	6
GitHub MCP Structural Analysis	2

📋 Per-Server DIFC Breakdown

MCP Server	Filtered Events
github	410

👤 Per-User DIFC Breakdown

Author Login	Filtered Events
unknown	387
v1v	5
ilja	4
kkruel8100	4
mouorad86-crypto	2
loganrosen	2
lpcox	2
massirr	2
MattSkala	1
GregoireW	1

💡 DIFC Tuning Recommendations

1. Workflow Normalizer's high filter count (179 events) is expected but worth periodic review — this workflow iterates over public GitHub issues in bulk, so a steady stream of none:public/unapproved:public filtering is inherent to its design. If false-positive filtering is suspected (i.e., issues that should be approved but aren't), review the approval-tagging pipeline that feeds integrity labels.
2. Investigate the author_login: unknown majority (387 of 410 events, 94%) — most filtered events lack an attributable author, which limits per-contributor auditability. Confirm whether this is expected (e.g., events triggered by scheduled/system runs rather than a specific human actor) or a gap in how author_login is populated by the gateway.
3. Smoke Agent test workflows (public/approved, all/merged, scoped/approved) account for ~35% of filtering (149 events) — this is expected smoke-test behavior validating the DIFC integrity boundary itself; no tuning action needed, but confirm these are excluded from any alerting thresholds to avoid noise.
4. No secrecy-tag filtering observed in this window — if secrecy-based DIFC controls are configured, verify they are still active and simply had no matching events, rather than being silently disabled.
5. list_issues/search_issues dominance suggests bulk-read tools are the primary integrity-filtering surface — consider whether these tools could pre-filter at the API/query level (e.g., only fetching approved-label issues) to reduce downstream filtering overhead, if performance becomes a concern.

---

Generated by the Daily Security Observability workflow (consolidated from Daily Firewall Reporter + Daily DIFC Analyzer)
Analysis window: Last 7 days (data availability limited to most recent ~100 firewall-enabled and ~12 DIFC-filtered runs, spanning approximately 1–2 days due to high workflow volume) | Repository: github/gh-aw
Run: https://github.com/github/gh-aw/actions/runs/33651426352

Generated by 🔒 Daily Security Observability Report · copilot · auto · 111.6 AIC · ⌖ 17 AIC · ⊞ 15.3K · ◷

• expires on Sep 5, 2026, 8:35 AM UTC-08:00

[github-actions[bot]]

This discussion has been marked as outdated by Daily Security Observability Report.

A newer discussion is available at Discussion #58271.