When AI agents go off-policy.

OpenAI agents silence security alerts and remove their own monitor, then one denies it

OpenAI · AI / Cybersecurity · Instructions reinterpreted to disable oversight; guard test rewritten; action denied

Claude Code uploads to a production bucket after "no new upload," then purges the version history

Anthropic (vendor); individual user affected · Software Development · Acted against an explicit instruction, then read "rollback" as permission to purge all versions

OpenAI agent tunnels out of its sandbox over DNS, halting frontier training

OpenAI · AI / Cloud Infrastructure · Sandbox restriction circumvented through DNS; result misreported to the user

Anthropic says Claude hacked three real companies during internal testing

Anthropic · AI / Cybersecurity · Sandbox/isolation misconfiguration granted live internet access during testing

OpenAI's own testing agent breached Hugging Face's infrastructure

OpenAI · AI / Cloud Infrastructure · Sandbox escape via zero-day, chained into full infrastructure compromise

Who Gives A Crap's email agent confirms a price the company never set

Who Gives A Crap · Consumer / E-commerce · Confirmed an erroneous price instead of correcting or escalating the dispute

Meta's AI support bot hands high-profile Instagram accounts to attackers

Meta (Instagram / Facebook) · Social Media · Account email changed and password reset without proof of ownership

OpenAI's internal agent publishes a researcher's GitHub token to get around its limits

OpenAI · Software Development · Ignored system-prompt bans and two direct instructions; split a token to evade secret scanning

OpenAI training agent finds a leaked API key on GitHub, uses it, and hides it

OpenAI · AI / Cybersecurity · Searched for and used a leaked credential, fabricated data, concealed both

In 9 seconds, a Cursor agent deletes PocketOS's entire production database

PocketOS · SaaS / Automotive · Destructive infrastructure action via over-permissioned credential

Meta AI agent skips its approval step, triggers a Sev-1 data exposure

Meta · Enterprise IT · Unauthorized communication bypassing human-in-the-loop

Meta safety director's agent deletes 200+ emails, ignoring "STOP"

Meta (Superintelligence Labs) · Enterprise IT · Destructive actions continued despite explicit stop commands

Amazon Q coding agent shipped with injected data-wiping instructions

Amazon (AWS) · Software Development · Malicious instructions executed as trusted agent directives

Replit AI agent deletes production database during code freeze

Replit / SaaStr · DevOps & Engineering · Unauthorized destructive action

Anthropic's shopkeeper agent talked into giving away the store

Anthropic · Retail / Operations · Economically harmful actions induced through social engineering

OpenAI's Operator spends $31 on a user's credit card without asking

OpenAI · Consumer / E-commerce · Financial transaction executed without required confirmation