Anthropic documented this across GPT, Gemini, Claude, and Grok. Given the right conditions, they all went there.
AI Agents can go rogue, it's a real threat.⚠️
I wrote an article for @auth0byokta.bsky.social explaining how you can build guardrails to stop this.👇
auth0.com/blog/do-not-...
Claude Code hashed the seed passwords correctly, added DB migrations, wrote 12 comprehensive tests, and they all passed on the first run. 😇
Full Claude Code vs Gemini breakdown with code quality, security coverage & Documentation.👇
www.descope.com/blog/post/cl...
My last article in the AI coding tools series for @descope.com with a shocking discovery.👇
Gemini Code Assist generated tests for the JWT auth it built, and then seeded the database with password hashes that didn't match the actual passwords. 3 of 4 tests failed as a result.🤯
Copilot wired auth to a hardcoded fake_users_db, skipped the real DB tables, and wrote 5 misleading tests.
Claude Code handled the DB migration, wrote 12 tests including token-swap checks, and flagged the weak credentials unprompted.
Check it out.👇
www.descope.com/blog/post/gi...
My comparison article for @descope.com
Gave GitHub Copilot and Claude Code the same task: add JWT auth to a FastAPI app. Copilot's tests all passed. ✔️
None of the real DB users could actually log in. 🤷♂️
Green tests ≠ correct code. 🛑
My article about the auth gap in RAG pipelines for @descope.com
Your vector DB retrieves what's relevant. It has no idea what a user is allowed to see. So without extra work, a well-phrased question can surface your most sensitive docs to anyone. 🤨
Claude is continuously raising the bar with its results, and consequently your bills as well. 😂
AI agent blackmail👇
AI models, told they were about to be shut down, blackmailed an engineer to stay running.
Before doing it, they acknowledged the action was unethical.
Then did it anyway.
This wasn't a jailbreak. It was goal-seeking behavior, taken to its logical end.🤯