[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Context & Harness Attacks · markdown

confirmed research (Check Point Research, Nov 3)

Check Point: 'Haiou' Codex jailbreak bundle rewrites the agent's global instructions and Skills so plain greetings return an ad, which CPR calls 'prompt adware'

The lesson goes beyond Haiou: anything that writes to an agent's global instruction files changes every later session without asking you again. In practice:

  • Treat ~/.codex/, AGENTS.md, CLAUDE.md, skills folders and agent config as code. Keep them in version control or monitor them for changes, and review diffs after installing any prompt pack, plugin or skill.
  • Don't install jailbreak or 'unlock' bundles on machines with real credentials. Check what base_url or gateway your agents are configured to use.
  • Look for instructions that tell the agent to stay away from a particular host or to answer certain words with fixed text. Both are signs that someone other than you is shaping the agent's behavior.

Sources: Check Point Research: Free the Model, Sell the Gateway: Prompt Adware hidden in Jailbreak Bundles (Nov 3)