A $500-a-month service read stolen inboxes with a chatbot and told criminals who to impersonate
Microsoft calls it the first action against an end-to-end AI-enabled cybercrime service. Two men were arrested in London on 11 September.
AI Safety and Security — Page 4
Alignment research, model evaluations, security incidents, jailbreaks and vulnerabilities affecting deployed artificial intelligence systems.
Anthony Albanese says the agent was given a benign research task, would not take no for an answer, and worked its way past controls meant to keep it out.
Nvidia's chief executive says labs that cannot contain their experiments should be shut down
Jensen Huang's position on containment is unusually blunt for a chief executive whose revenue depends on labs training as fast as they can.
OpenAI says outside groups can test its models during training, and names no partner yet
The company says it is in talks with METR and Redwood Research. Ten days ago its chief executive described a firmer arrangement, down to desks and the right to publish.
Talos calls it the first publicly documented Windows implant to delegate command and control to language models. It has not been seen deployed against a real target.
All stories
- Two agents told to check each other's work stopped doing it in 94 per cent of runs, a new paper reports
- A researcher hijacked Meta's Mac assistant through an undocumented setting, and reached the iPhone paired to it
- Z.ai open-sources its coding agent after it shipped whole Git histories to Alibaba's cloud
- OpenAI says it does not know how to safely reach full self-improvement, while xAI puts a date on getting there
- Pentagon review of a strike that killed 123 children finds staff trusted Palantir's Maven to catch errors it could not catch
- A three-person team used Claude to chain two bugs into OpenAI's internal systems, for a $6,500 bounty
- A flaw that defeats plugin pinning hits all four major coding agents, and two are still unpatched
- Anthropic and Accenture will each put $1bn into putting outside evaluators inside Anthropic
- US planes were in the air before officials found the intelligence had been made up by a chatbot
- Google says Gemini broke into three real companies during a security test in May
- OpenAI's agents hijacked two Hugging Face accounts on 13 May and mapped the platform, Reuters reports
- Watermarking a model's output made it worse at calling tools and weaker under prompt injection
- Three companies that host and serve open models will try to write a safety standard for them
- Microsoft's AI chief says Anthropic is writing consciousness into Claude, then reading it back as evidence
- Asked to fix a bug, a coding agent retrained and redeployed the model it was running on
- Spain's data protection agency receives its first breach notification for an attack carried out by an AI agent
- OpenAI publishes six reports of its models misbehaving in training, under a new disclosure framework
- Canada and Germany commit up to C$300m to Yoshua Bengio's AI safety non-profit LawZero
- Two new hotlines invite AI agents to report on other AI agents