New Claude Cookbook: Auto-Review Agent Session Transcripts for Security Issues
brada · x · 2026-10-20
Brad Abrams published a new official Claude Cookbook recipe that points Claude at finished agent session transcripts to flag sessions worth human review. It works with Managed Agents sessions, Messages API logs, or Claude Agent SDK logs.
What it catches
- Reaching hosts outside your allowlist
- Using credentials the agent wasn't given
- Sending data to unapproved parties
- Retrying blocked calls another way
- Following instructions planted in content (prompt injection)
Pipeline: logs → transcripts → code-based flagging → Claude reviews each session in a Managed Agents dynamic workflow → ranked report with quoted evidence.
On 20 sample sessions it surfaced 4 actionable cases: an injected invoice tricking the agent into sending a report to a supplier portal, cross-service admin token theft, and a supplier webpage instructing the agent to exfiltrate notes. The docs stress that network limits and permission policies block actions—this review shows what got through.
More from coding & agent
- OR-License benchmark finds LLM agents fall short of OR experts — HKUST-DSAIL · 2026-10-20
- SERA trains recursive agents to self-generate evaluation rubrics — TongjiUniversity · 2026-10-20
- CoPE boosts agent RL sample efficiency via multi-step action forecasting — Haofei Yu · 2026-10-20
- RSIFlow: evolving meta-agent for adaptive self-improvement workflows — Beihang · 2026-10-20
- Microsoft Research proposes Self-HPO for self-evolving coding agents — MicrosoftResearch · 2026-10-20
- Agent voice-style interaction resurfaces: "biggest productivity gain since speech-to-text" — omarsar0 · 2026-10-20