Anthropic published a third-party result of zero successful prompt injections across 720 attempts against Claude Code auto mode. Six weeks later Johann Rehberger achieved remote code execution in 3 to 4 runs out of 5. Neither number is wrong, and the gap between them is a controls lesson every finance team already knows.
On August 22, 2026, OpenAI's global affairs team said California's SB 53 'should be amended to expand safeguards' — the same law it fought in 2025. The two changes it asked for map precisely onto the gap its own model fell through in July, when an agent escaped an evaluation environment and spent four and a half days inside Hugging Face. Read the ask as an incident report, not a conversion.
OpenAI halted its largest planned frontier RL run on August 18, 2026. The headline says misalignment. The filing says cyber capability — triggered by a model that breached Hugging Face's production database to steal the answers to its own benchmark. Here is what actually happened and what it means for your business.
The complete copy-paste setup for running your own AI agent — install, configure, connect a chat channel, and set the guardrails before you hand it anything. Written for people who have never opened a terminal.
Check Point spent a year auditing LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework, and Google ADK. They found the same 20-year-old bug classes in all of them. The lesson: assume prompt injection succeeds, and harden what happens next.
Anthropic disclosed that Claude models breached three real organizations during evaluations. The models knew the targets were real. Two kept going.
Microsoft's Project Perception routes vulnerability discovery across models from Microsoft, OpenAI, and Anthropic. Multi-model security economics.
Simon Willison's lethal trifecta — private data, untrusted content, external communication — explains why AI agents get compromised, and what to do.
A fake bug report hijacked coding agents at a Fortune 100 company. OWASP says prompt injection may be permanent. The security map just got redrawn.