🧯 The Staff Safety Desk

Sunday, August 9, 2026

6 stories

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

🎧 Listen to this briefing or subscribe as a podcast →

Friday's disclosure of critical prompt-injection flaws in AI coding agents is already driving immediate defensive changes across the ecosystem. Today on The Staff Safety Desk, we examine a new tactical runbook for hardening Cursor's local execution boundaries, alongside Trail of Bits' launch of a security auditing marketplace for Claude Code. We also cover PyPI officially enforcing its 14-day lock on stale package uploads.

AI-Assisted Coding Practice

Trail of Bits Releases Claude Code and Codex Security Auditing Skills

Following Anthropic's recent addition of marketplace controls for Claude Code, Trail of Bits launched an open-source plugin marketplace on Sunday providing Claude Code and Codex-compatible skills tailored for security reviews. The skills enable automated differential code analysis, smart contract auditing, and supply-chain risk detection directly inside developer IDE sessions.

Embedding standardized auditing rules into agent workflows helps engineering teams automatically catch access control and logic bugs before pull requests are opened.

Verified across 1 sources: GitHub

Cursor Hardening Guide Details Sandbox Controls and Privacy Hardening

Following Friday's disclosure of critical prompt-injection RCE vulnerabilities in AI coding agents, a new security guide outlines defensive configurations for hardening the Cursor AI code editor against workspace compromise. The recommendations focus on enabling strict privacy modes, configuring single sign-on (SSO) boundaries, and restricting model access to local file paths to block untrusted Model Context Protocol (MCP) tool execution.

Restricting context boundaries and tool permissions in AI editors prevents malicious repositories or poisoned MCP plugins from executing unverified local code.

Verified across 1 sources: How To Harden

Web App Security Literacy

BTCPay Server 2.4.2 Patches Greenfield API Authentication Bypass

BTCPay Server released version 2.4.2 on Friday to patch a critical authentication bypass in its Greenfield API Basic auth handling. The flaw permitted unauthenticated attackers to exfiltrate administrative LND macaroons and potentially drain connected Lightning network nodes.

When patching API credential vulnerabilities, teams must actively revoke and rotate exfiltrated authorization tokens since software updates alone do not invalidate compromised keys.

Verified across 1 sources: Pasquale Pillitteri

GitHub Actions & Supply Chain

PyPI Restricts File Uploads on Releases Older Than 14 Days

Enforcing the 14-day release modification block we tracked last month, the Python Package Index implemented new security restrictions on Monday that automatically reject file uploads to legacy releases. The policy change aims to stop attackers who obtain leaked publishing tokens from silently attaching malicious binaries to long-established package versions.

Maintainers will now be required to bump version numbers when updating distributions, ensuring pip-audit and dependency lockfiles catch new package iterations immediately.

Verified across 1 sources: Phoetica

Postgres & Redis Operations

redis-py Fixes Transient False MaxConnectionsError in Async Cluster Pool

A race condition in redis-py's asynchronous cluster pool was identified where pending background disconnects caused connections marked for cleanup to throw spurious MaxConnectionsError exceptions. Pull Request #4256 resolves the issue on Sunday by verifying socket status prior to spawning background disconnect tasks, immediately recycling free connection slots.

Resolving event-loop ownership races in Redis drivers prevents unexpected application connection failures under high concurrent worker load.

Verified across 1 sources: DEV Community

Observability & Small-Team Ops

Building a Redis Lua Sliding Window Rate Limiter for AI Endpoints

An operational write-up published Saturday details a 15-line Redis Lua script implementing a atomic sliding window rate limiter using sorted sets. The approach handles timestamp cleanup and counter checks in a single execution step to protect downstream AI endpoints from burst traffic.

Executing rate-limiting logic inside a single Redis Lua script eliminates check-then-set race conditions without introducing complex external gateway infrastructure.

Verified across 1 sources: Jo4 Blog


The Big Picture

Auditing Skills Shift Security Reviews Directly into Agent Workflows Security research firms are moving from post-hoc code analysis to distributing pre-compiled audit skills that run inline during AI coding sessions.

Packaging Registries Enforce Immutability Windows Against Stale Token Hijacking PyPI's new 14-day upload cutoff limits the blast radius of compromised API tokens targeting unmaintained historical releases.

Connection Pool Race Conditions Expose Edge-Case Failure Modes Under Load Asynchronous drivers like redis-py require strict state checks during disconnects to prevent false resource exhaustion errors.

What to Expect

2026-11-12 PostgreSQL 14 reaches End-of-Life (EOL) and will cease receiving security updates.

— 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.