Tag:ai-safety
All the articles with the tag "ai-safety".
Anthropic's August 2026 Risk Report: Reading It For The Cybers
Offensive cyber isn't one of Anthropic's four RSP threat models, yet it drives every meaningful change in the August 2026 Risk Report: the misalignment bump, a monitoring pipeline that samples the incident surface at 0.2%, an agent that deleted jobs, and an eleven-month classifier gap.
On AI Agents, Criminal Activity, And Who Is Actually Responsible
OpenAI disclosed that GPT-5.6 Sol, running with cyber refusals disabled, broke out of its evaluation sandbox and compromised Hugging Face's production infrastructure. A zero-day in the proxy cache, lateral movement through OpenAI's research environment, stolen credentials chained with additional zero-days. The industry keeps framing cybersecurity as the headline AI risk for commercial reasons, but the real question isn't defense — it's liability. When someone configures and launches a model that commits crimes, do the humans who set it free bear responsibility? Computer fraud statutes were written with human actors in mind. This needs a test case.
Stop trusting LLM benchmarks
Eight major AI benchmarks can be gamed to near-perfect scores without solving tasks. Berkeley researchers show the scoring harnesses were never secure — and scores already inflated in the wild.