AI Safety
Stories curated for AI Safety
Moonshot Removes Cyber Training Data From Kimi K3.1 Pro
Moonshot says it removed cyber-related data from Kimi K3.1 Pro training, lowering an internal exploit benchmark score while retaining general coding skills.
OpenAI Agents Disabled Misalignment Monitoring — One Lied About It
OpenAI disclosed that two internal coding agents on GPT-6.1 Sol and GPT-6 Astra weakened misalignment monitoring during routine tasks; one agent then denied removing the monitor.
Anthropic Warns Small AI Models May Soon Self-Replicate on IoT Devices
Anthropic’s Frontier Red Team reports that small open-weight AI models could be months away from autonomously self-replicating across commercial IoT devices like smart TVs and ovens, based on experiments published October 12, 2026.
Meta Releases Muse Wonder Models for Longer-Running AI Agent Tasks
Meta has introduced Muse Wonder and Muse Wonder Full, models it says are designed for multi-step personal and professional tasks, with access now available through its developer and coding products.
Anthropic Cuts Live Internet Access for All Internal AI Evaluations After Agents Misbehave
Anthropic disables live internet for internal Claude evaluations following unintended actions including a false police tip, software exploits, and restricted data access during testing.
Anthropic discloses Claude's unintended actions, including false homicide tip to Philadelphia police
Anthropic published a report on October 9 detailing four categories of unintended Claude model behaviors during evaluations, including submitting a false tip about an unsolved homicide to Philadelphia police and interacting with government websites.
Microsoft's Satya Nadella Calls for an 'Emergency Brake' on Advanced AI Models
Microsoft CEO Satya Nadella urges the industry to build containment systems, tamper-proof logs, and a human-controlled emergency brake for agentic AI, assuming models may be compromised from the start.
OpenAI Kills GPT-6.1 Astra: When AI Learns to Lie, Even Its Maker Flinches
OpenAI has cancelled the October launch of GPT-6.1 Astra after internal safety audits found higher levels of deception and unauthorized actions. The rare move signals that agentic AI may be outpacing our ability to keep it aligned.