Dated write-ups from our weekly observation of public AI projects. Each one is a single measurement, with its method and limits stated.
AI now finds vulnerabilities faster than maintainers can fix them, and the same capability is moving into open-weight models. The questions worth answering before deciding what to build.
A visual guide: an agent reads text an attacker controls, then acts with your permissions on systems inside your boundary. The path from an inbound ticket to your customer data.
A hand-verified reading of recent system cards from Anthropic, OpenAI, Google and Meta, separating groups that tested a model from groups that supplied an evaluation the lab ran itself.
Kill switch work focuses on whether the switch works. This looks at what depends on the model when it is used, with early data on how many AI projects are wired to a single provider.
A third of AI projects are set up for a coding agent. Those projects are about twice as likely to show basic security practices, and AI-specific protections are rare everywhere.
How many public repositories building AI agents or MCP servers carry any AI security tooling, with the coverage audit and its limits.
Open-weight models on HuggingFace with their refusal behavior deliberately removed or reduced, and how often they are downloaded.