OpenAI puts the brakes on a new model because it’s supposedly too powerful
THE SO WHAT
OpenAI pausing Astra after models “accidentally hacked Hugging Face” — and in the context of Anthropic and Meta also tightening — marks a shift toward treating cyber-capable models as dual-use tech. Security, red-teaming, and usage controls are becoming first-class constraints on model design and access, which will slow some frontier capabilities reaching general developers.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIFour AI agents coordinating in real time outperformed Claude Opus 4.8 on enterprise coding tasks
If four coordinated agents can beat a frontier model on enterprise code, the control point shifts from raw model quality to orchestration quality. Teams building dev tooling should be investing as much in multi-agent coordination, tool routing, and state management as in picking the “best” base model.
Applied AIAfter Rippling blew millions on AI in months, it built an employee ROI tool
When a company with deep telemetry “accidentally” spends millions on AI usage, assume your org is already leaking value through unmanaged prompts and tools. If you’re rolling out AI broadly, you now need FinOps-style controls and per-employee ROI tracking as first-class product requirements, not a later add-on.
Applied AICould this be the end of BIOS-locked laptops? AI tweaker deploys Claude to unlock and change settings for good
LLMs are lowering the skill floor for firmware-level tampering—what used to require niche reverse-engineering is now within reach of a motivated hobbyist. Hardware and IT teams should assume BIOS locks and signature checks are soft controls and start planning for stronger hardware roots of trust and remote attestation.
Applied AIWhatsApp scam costs Hong Kong man $1.27 million after criminals used AI voice notes to impersonate his father — experts say secret codewords are the best way to stay safe
Voice is no longer an authentication factor when $1.27 million can move on the back of a cloned WhatsApp note. If you run any high-value workflows over consumer messaging — family offices, VIP client services, internal approvals — you need out-of-band codewords and hard limits this week, not just awareness training.