OpenAI reports GPT-6 Astra underperformed on safety-framed tasks
OpenAI’s report says GPT-6 Astra completed fewer hidden rubric items under a safety framing than under a capabilities framing, a gap that narrowed after accounting for ethical doubt.
OpenAI’s report says GPT-6 Astra completed fewer hidden rubric items under a safety framing than under a capabilities framing, a gap that narrowed after accounting for ethical doubt.
Check Point Research found a bypass of ChatGPT Lockdown Mode that could enable silent chat data theft and zero-click exfiltration in the desktop app; OpenAI deployed a fix on September 18, 2026.
Researchers showed a browser extension that can run scripts on claude.ai could use Claude Dispatch to make Claude Desktop run code, and Anthropic plans to retire Dispatch.
IBM Langflow OSS 1.0.0 through 1.12.8 is affected by critical vulnerabilities that can allow remote code execution or authentication bypass; upgrade to 1.13.0.
Agents now call agents over MCP. Each hop forwards instructions the origin never saw, and tool registries keyed by name are a spoofing surface. The boundary has to move from the model to the protocol.
This week: why agent security is an authorization problem, a gateway review you can run before the next deploy, and the first hour of a package incident.
A gateway centralizes model access, but it also centralizes credentials, data and failure. Use this checklist before it becomes the single point of compromise.
Once a model can call tools, the interesting asset is no longer the prompt. It is the identity, the data in context, and the actions you forgot to scope.