ai-security
22 posts
Most automation doesn't need a smart model
How 8-29MB automation models like Cactus Needle 3 match large models on narrow tasks, where they fail, and the security tradeoffs of running them locally.
The hallucination was not the failure.
An AI-generated intelligence report reached a US Military context and caused a close call because generative output was trusted as verified intelligence.
The model already broke your embargo
An OpenAI model reproduced the RubyGems caching vulnerability before public disclosure, exposing why embargoes fail against automated ingestion.
Three filled lines in the HuggingFace postmortem
A HuggingFace hack postmortem by METR and Redwood confirms only authorship, subject, and focus. Mechanism is not confirmed, so no control change is authorized.
Grok packed your home folder into an API request
How AI assistants with filesystem access end up transmitting your home directory to vendor servers - and the concrete steps to scope, verify, and contain it.
Mythos AI cleared for distribution, no validation report
REDLINE breaks down the security risk in releasing Mythos AI to trusted US organizations: not the model, the missing adversarial validation and zero prompt-level telemetry.
A SQLite underflow, and the flood behind it
AI isn't replacing defenders - it's multiplying vulnerability volume, hallucinated dependencies, and synthetic findings. The skill that survives is validation at machine rate.
speed ships the flaw
AI generates code, config, and infrastructure faster than validation can check it. The gap between production rate and validation rate is the attack surface.
The same AI you're shipping wrote the malware
10,000 trojan GitHub repos weren't a malware breakthrough - they prove LLM safety lives in the model while abuse happens in the unguarded pipeline.
The contract you pasted is now giving orders
Large AI context windows turn conversations into unsecured databases, breaking DLP assumptions and opening prompt injection paths. Here's how to reassess the risk.
Typosquatted Microsoft AI packages harvest developer credentials
How attackers weaponised typosquatted Microsoft AI tooling to harvest OpenAI, HuggingFace, AWS, and Azure credentials from developer workstations.
Meta's chatbot worked exactly as designed.
Meta's AI chatbot enabled mass Instagram account takeover by resolving conversational framing into identity actions through sanctioned internal workflows.