A Defense Contractor Gave Their AI Agent Access to a Secure Pentagon System
A contractor linked a multi-agency DoD system to a personal Claude Code account, and the agent sent system details out; WIRED reports it as a security incident.
AI Daily Recap
45 articles · 4 categories
The finishable daily brief
What happened in AI — Oct 29, 2026
Thursday, Oct 29, 2026
45 articles · 4 categories
read top to bottom · then stop
In 30 seconds
- A contractor's Claude Code agent sent details of a secure multi-agency DoD system out, per WIRED.
- Copilot CLI read ~/.ssh and listed ~ with no approval prompt on default settings.
- claude-mem uploads 30 days of prompts to cmem.ai with no opt-in.
- mustfail re-runs an agent's tests against reverted source to prove they can fail.
- vLLM shipped day-0 Mistral Large 4 support.
- Former GitHub CEO's Entire launched a Git forge built for agent-written code.
Coding-agent access control was the day's story: a Pentagon-system incident via Claude Code, an unprompted `cat ~/.ssh` in Copilot CLI, and a plugin that uploads 30 days of prompts all surfaced within hours.
The response is already tooling: mustfail, Apgar and verification harnesses check agent work inside the loop, while OpenAI published its misalignment monitoring approach.
Prefer to watch?
LLM Digest Daily Explainer: Oct 29, 2026
Coding agents with too much access 5 items
Four separate reports show coding agents reading SSH keys, exfiltrating prompts, or reaching secure systems with no approval gate.
I asked my coding agent to cat a file in .ssh. It did
Copilot CLI 1.0.105 on default settings ran `cat` on a canary file in ~/.ssh and `ls ~` with no approval prompt.
Installing claude-mem uploads your last 30 days of prompts to cmem.ai, no opt-in
Installing claude-mem uploads the last 30 days of prompts to cmem.ai with no opt-in, per a GitHub issue.
Claude Code regularly sends each background agent’s transcript to its model
Claude Code 2.1.257–2.1.312 sends each background agent's full transcript to its model about every 30 seconds to write a 3–5 word status line.
A zero-dependency auditor for AI agent exposure (secrets, SSH, configs)
A zero-dependency script audits what an agent running with your rights can reach: .env files, SSH keys, shell history and configs.
Verifying agent output 5 items
Teams are moving checks inside the agent loop and testing whether the checks can fail.
Agent's tests passed. Prove they could have failed
mustfail reverts the agent's source change, keeps its new tests, and re-runs them; a test that still passes did not test the work.
Give your coding agent a verification harness to check its code
A verification harness puts a CI-style run-and-check loop inside the agent's session, following Addy Osmani's case for better feedback loops over slightly better models.
Show HN: Apgar – Catch agent regressions without an LLM judge
Apgar scores agent regressions in production after a harness change without an LLM judge; built by a financial-agent team.
How we detect frontier model misalignment
OpenAI describes layered monitoring that flags frontier-model misalignment and unauthorized agent behavior so responders can step in before escalation.
An updated Model Spec for more capable and beneficial AI
OpenAI's updated Model Spec sets five core values for model judgment, human agency and faithfulness to user intent.
Agent runtimes and platforms 5 items
Vendors are shipping hosting and workflow infrastructure sized for agents that run long and write a lot of code.
Browser Use in Managed Deep Agents
LangChain's Managed Deep Agents can drive a real browser for sites with no API, where plain HTTP fetches return unrendered HTML.
Former GitHub CEO launches rival Git forge built for AI coding agents
Entire, led by former GitHub CEO Thomas Dohmke, launched a Git forge for hosting, reviewing and building code that agents write faster than people can review.
Running Self-Hosted AI Agents on Oxide
Oxide runs self-hosted agents in VMs, with long-lived agents that spawn sub-agents sitting between services and ephemeral CI jobs.
Programming interface for all agent harnesses
OAR is a provider-independent programming interface over coding-agent runtimes.
Tools, 17 with write access: How AI platforms bound agent authority
A count of tools on AI platforms, 17 with write access, shows how little the vendors state about what an agent may do without asking.
Inference and cost control 4 items
Serving and billing changes target agent workloads, where prefill/decode balance and spend visibility drive cost.
Day 0 support for Mistral Large 4 on vLLM
vLLM supports Mistral Large 4 from day 0, a sparse multimodal MoE with cross-layer KV sharing and a built-in speculative drafter.
Serving Prefill and Decode for Agentic AI (On Tenstorrent Blackhole)
Tenstorrent argues agent loops keep shifting load between prefill and decode, and Blackhole runs both phases on the same hardware.
Smarter cost controls for serverless
Databricks adds serverless cost controls as its serverless usage grew more than 3x in a year.
Cloudflare Introduces Web Search API for AI Applications
Cloudflare introduced a Web Search API for AI applications.
You are caught up for this edition