This release includes 1 security fix for security teams reviewing exposed deployments.
Topics
+14 more
Affected surfaces
ReleasePort's take
Moderate signalThe v5.37.3 release fixes path‑approval to enforce unconfined resolveBase and adds a read‑denylist floor; it also introduces best‑effort root path scanning for bash security hardening.
Why it matters: Security hardening now scans root paths in bash (severity 90) and enforces denial lists, directly reducing attack surface for deployments using these components.
Summary
AI summarySecurity hardening adds best‑effort root path scanning and denial‑list enforcement for bash.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Security | Critical |
Adds best‑effort readRoots/writeRoots path scan for bash security hardening. Adds best‑effort readRoots/writeRoots path scan for bash security hardening. Source: llm_adapter@2026-07-14 Confidence: high |
— |
| Bugfix | Medium |
Fixes path‑approval to honor unconfined resolveBase and adds read‑denylist floor. Fixes path‑approval to honor unconfined resolveBase and adds read‑denylist floor. Source: llm_adapter@2026-07-14 Confidence: high |
— |
Full changelog
What's Changed
- test(tui): real-PTY scrollback harness for the compositor (#541) by @griffinwork40 in https://github.com/griffinwork40/agent-afk/pull/553
- fix(path-approval): honor unconfined resolveBase in the hook + add read-denylist floor by @griffinwork40 in https://github.com/griffinwork40/agent-afk/pull/560
- security(bash): best-effort readRoots/writeRoots path scan + ADR (closes #354) by @griffinwork40 in https://github.com/griffinwork40/agent-afk/pull/558
Full Changelog: https://github.com/griffinwork40/agent-afk/compare/v5.37.2...v5.37.3
Security Fixes
- security(bash): best‑effort readRoots/writeRoots path scan + ADR (closes #354)
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Agent AFK
Coding-agent harness built for unattended, away-from-keyboard runs across four surfaces sharing one session manager: one-shot CLI, REPL, cron-friendly headless daemon, and a Telegram bot. Editable agent loop (prompts, permission gates, model routing, explicit terminal states), MCP (stdio/HTTP/SSE + OAuth), lifecycle hooks, background subagents, plan mode, cross-session memory, and an append-only trace receipt (`afk trace show`). Works with Anthropic and any OpenAI-compatible endpoint (GPT, Codex, local MLX/llama.cpp/Ollama)
Related context
Related tools
Beta — feedback welcome: [email protected]