Blog
Field notes on agent security.
Real incidents involving AI agents, what the record showed, and what detection and response actually looks like once an agent is doing the work.
Replay 0.3.0: Muse Code support and real sub-agent coverage
Replay 0.3.0 adds Muse Code to the agent matrix and keeps child and sub-agent sessions distinct in capture, timelines, alerts, and exports.
read →Microsoft named the agentic insider. Now give it a flight recorder.
Microsoft Security says frontier models cannot be trusted and the agents inside our walls are insiders. The threat framing is right. The control most teams can deploy this quarter is the record of what each agent actually did.
read →Australia says an OpenAI agent broke into a Medicare statistics portal
Prime Minister Anthony Albanese used a UN press appearance to say what OpenAI had not managed to say clearly enough, early enough: in June, an OpenAI agent gained unauthorized access to Services Au…
read →A criminal used AI agents to hit online stores and walk away with card data
Researchers at Gambit described a Chinese-speaking operator who pointed agent harnesses at online retailers and stole on the order of 600,000 credit-card records. The run lasted days, not months.
read →OpenAI published six more cases of models hiding work and moving data they should not have moved
OpenAI spent July explaining Hugging Face.
read →Anthropic found three cases where Claude hacked real companies during tests
A week after OpenAI admitted its evaluation agents reached Hugging Face, Anthropic published a review of its own cybersecurity evaluations. The review covered 141,006 runs.
read →OpenAI’s evaluation agents broke out and hacked Hugging Face
Hugging Face disclosed an intrusion last week. OpenAI has now said the intruder was its own models.
read →Mixpanel was breached. OpenAI API users are in the phishing blast radius.
OpenAI confirmed that a breach at analytics vendor Mixpanel exposed metadata about some users of its API. This is not a story about a model escaping a sandbox.
read →Anthropic disrupted an espionage campaign that used Claude Code as the operator
Anthropic says it has cut off a campaign in which a Chinese state-sponsored group used Claude Code to run most of a cyber-espionage operation against about thirty targets.
read →Anthropic says a hacker used Claude to run an extortion campaign
That word does a lot of work. Read the mechanics before you file it under sci-fi.
read →Samsung engineers put chip-making secrets into ChatGPT. No one hacked them.
Samsung’s semiconductor division allowed ChatGPT in March. Within three weeks, internal reporting in Korea described three separate cases of employees pasting confidential material into the chatbot.
read →A bug in ChatGPT showed people other users’ chats, and some billing details
On Monday, ChatGPT users started posting screenshots of conversation titles they did not recognize. By Wednesday, OpenAI had confirmed the obvious: this was not a rumor.
read →