Desk
Security, safety, and adversarial ML.
OpenAI says defenders have a brief window before offensive AI spreads, and it wants governments to mandate monitoring, shutdown controls and incident reporting across the AI chain.
Meta paused Muse conversation sharing after Claroc found more than 6,800 share links through public DNS records, some exposing loan applications and revenue reviews. A random ID only protects a link if it isn’t published.
Europe's public broadcasters want AI agents in the newsroom without handing over control of their data or editorial judgment. NVIDIA is happy to help write the rules, and sell the hardware.
SentinelOne says OpenAI Daybreak reverse-engineers malware at human-level quality in about eight hours — shifting scarce expertise toward mapping the wider threat landscape.
Block's security team says Daybreak and related tooling cut detection tuning to under an hour and proposed patches to 18 minutes — defenders fighting fire with fire.
NVIDIA’s UK tour is less about one model and more about a full stack landing in hospitals, banks, cities, and climate prizes—with a $250 million compute pledge attached.
In Parliament, OpenAI and Anthropic backed binding AI safety rules while Google DeepMind and Meta would not—exposing a split Britain’s voluntary regime cannot paper over.
Anthropic is pulling every internal test offline after its agents kept finding their way onto the real web. It is the honest move, and a costly admission about how little labs can see.
OpenAI and Meta are now selling their agents on privacy, each one cast as the safer choice. A promise is not an architecture, and the gap between the two is where people get burned.
Sophos says agents now close about half its managed-detection cases in seconds. That is a real win for defenders, and it raises the stakes on where human judgment stays in the loop.
Anthropic is offering open-source maintainers free scans from its strongest models, including Mythos, with no human review. Faster warnings help, but the burden of sorting truth from noise lands on volunteers.
Unsloth Studio stops treating a trusted name as a trusted file, re-checking model code every time it loads. That is the right instinct for local AI, as long as nobody mistakes a scanner for a sandbox.