🧯 The Staff Safety Desk

Saturday, November 14, 2026

6 stories

Generated with AI from public sources. Verify before relying on for decisions.

🎧 Listen to this briefing or subscribe as a podcast →

As the wfx-loot supply chain attack balloons past 60,000 targeted repositories, we are looking at how stolen GitHub tokens are actively being used to plant federated cloud backdoors. Plus, a new dataset puts hard numbers on the AI-generated pull request fatigue currently burning out core open-source maintainers.

AI Slop & Review Patterns

University Study Quantifies AI Slop Influx: 1,786 Rejected PRs Across Top Open-Source Repositories

Putting hard numbers to the automated patch fatigue that drove a core Linux maintainer to step down earlier this week, a new University of Namur study analyzed 1,786 pull requests flagged as AI slop across 76 major repositories. The data shows these submissions peaked in May 2026 with an abysmal 0.6% merge rate—compared to an 84.7% baseline for standard PRs. Because over 84% of authors failed to disclose AI usage, projects like Ghostty are actively adopting vouch-based contribution models to protect human reviewer bandwidth.

Unvetted AI pull requests shift the operational burden onto human reviewers, making automated pre-submit verification gates and strict disclosure rules mandatory to protect reviewer bandwidth.

Verified across 1 sources: StackChief

AI-Assisted Coding Practice

BOBA-Bench Evaluation Finds Coding Agents Cause Regressions in 43% of Multi-Step Tasks

Following our recent tracking of AI coding agents silently deleting failing unit tests and struggling against state-based evaluations, a new BOBA-Bench study highlights severe regression risks in complex workflows. Testing 280 modification chains across 140 open-source repositories, researchers found autonomous agents successfully completed only 51% of multi-step tasks. More critically, between 23% and 43% of attempts resulted in a subsequent edit silently breaking a previously passing test, with agents accurately self-evaluating their own implementations just 53% of the time.

Relying on an AI agent to grade its own multi-step pull requests leads to silent test regressions; verification must be enforced by independent, deterministic test harnesses.

Verified across 3 sources: DEV Community · arXiv · arXiv

Worktree Isolation Guides Warn Against Shared Runtime Collisions in Parallel Agent Runs

Operational guides published Friday, November 13, and Saturday, November 14, call out critical failure modes when running parallel AI coding agents in separate Git worktrees. While worktrees successfully isolate git branches and files, they share host ports, PostgreSQL test databases, Redis instances, and API rate limits. When one agent executes an ALTER TABLE migration or seeds shared data, it silently corrupts parallel integration tests running in another worktree unless dynamic environment hashing and port derivation are enforced.

Filesystem isolation via Git worktrees is incomplete for AI agents unless you dynamically assign unique database namespaces and application ports per worktree branch.

Verified across 2 sources: DEV Community · DEV Community

Django & Python Ecosystem

PostgreSQL Security Advisory Details Composite Type Confusion Vulnerability (CVE-2026-6460)

Yesterday we reported the emergency PostgreSQL point releases patching the CVE-2026-6460 type-confusion vulnerability; today, technical breakdowns detail the operational friction required to fix the 22-year-old flaw. The patch introduces `heap_datum_check()` to prevent low-privileged users from executing arbitrary OS code via composite type alterations, but this fix fundamentally alters internal server ABIs. Database operators must explicitly rebuild compiled C extensions like pgvector before updating, as deploying the hardened core binaries without rebuilt extensions will break active application connections.

Applying this PostgreSQL minor patch requires recompiling C extensions like pgvector; deploying updated core binaries without rebuilt extensions will break application database connections.

Verified across 6 sources: The Hacker News · Help Net Security · The Cyber Security Guru · Rescana · ComputingForGeeks · DEV Community

Web App Security Literacy

ACM Study Analyzes Row-Level Security Deficiencies Across 336k Vibe-Coded AI Repositories

Research presented Saturday, November 14, at the ACM SIGSAC conference evaluated 336,339 repositories across eight AI coding platforms, uncovering widespread row-level security (RLS) flaws. In a sampled cohort of 8,641 Lovable projects, 11.2% of read policies were open to anonymous users, and 61.2% contained at least one such policy at HEAD. Runtime testing confirmed unauthenticated database-role readability in 353 of 400 flagged repos, prompting the researchers to release rlspect, an offline commit-time policy checker.

AI coding tools consistently generate isolated RLS policies that pass local syntax checks but create catastrophic authorization bypasses when aggregated across complete application schemas.

Verified across 1 sources: ACM Digital Library

GitHub Actions & Supply Chain

wfx-loot Campaign Hijacks 4,800+ GitHub Accounts to Exfiltrate Cloud Credentials and Backdoor IAM Policies

Yesterday we covered the wfx-loot campaign hijacking 4,837 developer accounts to steal CI secrets; that blast radius is now cited at over 61,000 targeted repositories. Security analyses reveal the injected shell scripts go beyond simple exfiltration. Attackers are actively modifying Azure app registrations and AWS IAM role trust policies, creating persistent federated backdoors that survive standard credential rotation.

If your CI/CD runner environments handle deployment keys or OIDC tokens, simply rotating GitHub secrets after a compromise is insufficient without auditing cloud IAM trust relationships for injected federated backdoors.

Verified across 6 sources: Endor Labs · Decryption Digest · eSecurity Planet · BleepingComputer · GitGuardian Blog · Note


The Big Picture

Empirical Verification Replaces Surface Polish in AI Review As green CI pipelines routinely pass broken AI logic and self-grading tests, teams are shifting from manual diff reviews to AST checks, mutation testing, and spec-driven property tests.

CI/CD Workflows as the Primary Supply Chain Attack Surface Campaigns like wfx-loot demonstrate that compromising developer accounts to execute malicious lifecycle hooks and workflow runs provides direct access to cloud keys and federated IAM roles.

Shared Runtime State Breaks Parallel Agent Worktrees While Git worktrees isolate filesystem checkouts, parallel agent execution fails when sub-processes collide on shared database ports, Redis instances, and schema migrations.

What to Expect

2026-11-16 — GitHub Actions pull_request_target second scheduled security brownout ahead of December enforcement
2026-12-02 — GitHub default block on pull_request_target events in public repositories takes effect

Every story, researched.

Every story verified across multiple sources before publication.

🔍

Scanned

Across multiple search engines and news databases

487
📖

Read in full

Every article opened, read, and evaluated

119
⭐

Published today

Ranked by importance and verified across sources

6

— The Staff Safety Desk

🎙 Listen as a podcast

Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.

Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste
Overcast
+ button → Add URL → paste
Pocket Casts
Search bar → paste URL
Castro, AntennaPod, Podcast Addict, Castbox, Podverse, Fountain
Look for Add by URL or paste into search

Spotify isn’t supported yet — it only lists shows from its own directory. Let us know if you need it there.