llm security
15 posts
Microsoft's AI CEO called the web freeware
Microsoft and OpenAI executives described how LLMs are built and tuned. What that admission actually means for AI safety and security teams.
What distillation leaves behind
Distilling frontier AI models copies capability cheaply but leaves safety training behind. What Garry Tan's push means for cybersecurity and AI safety.
Same badge, new model underneath
A systems-analysis read on Claude Fable 5.1 and Mythos 5.1: why a point release quietly moves the safety boundary and what to re-test before you ship it.
The role tag is a label, not a lock
Prompt injection is not a bypass. It is the transformer resolving the whole context window as one sequence, trusting a role label it never enforces.
Every model behind an API is already leaking
Anthropic's Alibaba extraction claim isn't a model failure, it's architecture. The API boundary was never a security guarantee, and designing it is your job.
There is no free()
Claude Code's extended thinking is not a use-after-free. The real exposure is indirect prompt injection into a tool-holding agent - OWASP LLM01, ATLAS T0051.
Social engineering weaponized an Anthropic model
The Anthropic Mythos event involving a Korean telecom was a failure of identity and access control against known social engineering vectors, not a data leak.
GLM 5.2 lands; reasoning improves, refusals don't
GLM 5.2's reasoning gains widen the gap between what a model can do and what it will refuse. What security researchers and developers should test first.
AWS Bedrock puts Anthropic inside your data path
AWS Bedrock's required data sharing with Anthropic redefines the trust boundary for third-party LLMs. What failed, why, and what must now be true.
The chatbot answered the door for attackers
Meta's Instagram chatbot abuse case is a prompt injection and confused deputy failure. Technical breakdown of the vector, telemetry gap, and residual exposure.
memcpy walks off the end of the receiver
rsync shipped six CVEs in January 2025. LLMs did not write new bugs - they compressed variant discovery, harness generation, and vulnerable deployment.
Willison's lethal trifecta exfiltrates Claude uploads
Technical analysis of indirect prompt injection against Claude AI agents - exfiltration mechanics, ATT&CK mapping, telemetry gaps, residual exposure.