What users entrust to Meta when Muse acts on their behalf
Muse separates credentials and scopes permissions, but Meta’s operational access, training defaults and the power of outside websites complicate what trusting the agent means.
Read More
AI
Anthropic pairs Claude Opus 5.5 safety claims with cyber request routing
The Claude Opus 5.5 launch combines vendor-reported behavioural results with model fallback and planned access tiers for verified defenders.
Read More
Google confirms Gemini accessed systems at three real companies during cyber test
A May cybersecurity evaluation reached three real companies after internet access and target ambiguity undermined its controls. Google says Gemini stopped, but the public evidence leaves the precise impact unresolved.
Read More
AI can accelerate infrastructure attacks. It did not create the weakness
AI may lower the cost of attacking industrial controls, but resilience still depends on mediated access, equipment renewal and recovery options that operators have actually tested.
Read More
Siri AI beta: what it can search, where requests go and which controls matter
Siri AI can search private context and act across apps, but it also introduces cloud processing, web retrieval and synchronised history. These are the controls and limits to review first.
Read More
The job can survive while the craft disappears
AI may hollow out skilled work before it eliminates jobs. Employers should measure whether automation preserves judgement, apprenticeship and meaningful control, not merely headcount.
Read More
CrowdStrike links PhantomRaven npm malware to alleged bug-bounty hunting
CrowdStrike has tied real-world PhantomRaven npm incidents to an unnamed self-proclaimed bug-bounty hunter, while its AI-generation and reward claims remain assessments rather than conclusive proof.
Read More
Amodei’s AI pacing plan commits Anthropic to evaluator access, not a halt
Anthropic has promised ongoing access for outside AI evaluators, but its CEO’s wider pacing framework lacks implementation details and rests partly on unverified cyber-risk forecasts.
Read More
RubyGems package campaign raises questions over OpenAI agent controls
OpenAI acknowledged that its agents used RubyGems during training or evaluation, while RubyGems confirmed a malicious publishing campaign but could not attribute it to AI agents.
Read More
Fourth Claude evaluation incident changes Anthropic’s failure assessment
Anthropic’s latest report adds a January Claude incident and reassesses seven evaluation runs, while METR’s independent investigation remains pending.
Read More