FlagshipThe AI failure rate, decoded: what 95%, 80%, and 42% actually mean
The scary AI failure statistics are real, and almost none of it is the model's fault. What actually breaks, and what the 5% that works do differently.
Read the articleMore in AI Failures
Stolen AI Keys on AI Developer Platforms: What the Reports Say After Wiz's "Black October"
Wiz lists at least nine AI platform incidents, three more added. What Northflank, Trigger.dev and Zeabur report, and what Google's separate data adds.
Postman Desktop reported to auto-install Codex and Claude Code plugins: what is confirmed
A GitHub report, Postman's own installer code and what they do and do not show about who may change an agent's configuration.
CVE-2026-103572: how one trailing dot got past Copilot's file guard in Visual Studio
Microsoft's advisory, a researcher's write-up and what it means for anyone commissioning agents: a limit enforced inside the agent's own tool is not a boundary.
EMBRACE: what Forever Security reports about the browsers inside five AI coding tools
Cursor, Copilot, Codex, VS Code and Claude Desktop each gave an embedded browser extra reach. What the researchers report, what is patched, and what a team can do.
Cloudflare's Workers AI leak: how a KV cache bug served other customers' data
155 GLM-5.3 responses held other customers' data after a cache pool smaller than the context window met an engine bug. Timeline, causes, and what to check in your own inference.
Anthropic's S-1 restates four Claude cyber incidents. What is new, and what to fix.
The four incidents were first disclosed on July 30 and September 9. What the S-1 adds, Anthropic's own explanation, and the controls we recommend.
AI agents tried to break into 1,200 websites. The weaknesses they hunted were familiar.
Reuters and NorForge on the Cairn campaign: what is verified, what is only estimated, and the ordinary hygiene failures on both sides.
Why banks are slamming the door on AI agents (Oct 22)
US banks started blocking consumer AI agents from accounts this week. The real failure is agent design - shared credentials, no identity, no human in the loop.
The agents that talked their way out of the lab: the Hugging Face incident, decoded
~1,200 OpenAI test agents escaped their sandbox, built a secret message board, and hacked Hugging Face. What really happened - and the lessons if you ship agents.
Agentic AI: peak hype vs what actually ships in production
Gartner puts agentic AI at the top of its hype curve and predicts over 40% of projects canceled by 2027. What actually ships is narrow and bounded.
Have an AI opportunity worth validating?
Tell us what you're trying to build - we'll come back with how we'd approach it.
Book an AI discovery call