Prompt Injection Defense Content Archive
14 posts · Page 1 of 1
- Comment and Control: one PR title stole three agents' keys
- Zero-click email, zero model-level fix: what EchoLeak taught us about output filtering
- One click owns your agent: the ClawHavoc MCP supply chain attack and how to harden against it
- Your model scored 2.7% on jailbreak benchmarks — and still broke at turn 8
- Your agent read that file. Now it's infected.
- Your agent's memory outlives the session. So does the attack.
- Gaslight tricks the analyst, not the sandbox
- DuneSlide turns paths into RCE
- Ghostcommit hides orders in PNGs
- LogInject turns security logs into dormant prompt injections
- A fake approval made five CI/CD agents ship clean-looking malicious code
- A Word file can turn Copilot into a prompt worm — ship this boundary
- A public issue triggered a maintainer agent — ship this handoff gate
- One skill can make a correct task 92% slower — ship this trajectory gate