#AI security

NHS England is facing criticism after directing staff to remove existing and future open-source software from public view, a response tied to concerns about AI systems that can find software flaws.

Anthropic says it is investigating a report that unauthorized users accessed its Claude Mythos Preview cybersecurity tool through a third-party vendor environment, while saying it has not found evidence its own systems/N

Mozilla says early access to Anthropic’s Mythos Preview helped identify 271 security vulnerabilities in Firefox 150, intensifying the debate over whether advanced AI will shift the balance between defen

Anthropic’s latest model, Mythos, is being made available only to a limited group of organizations after internal and external testing suggested it can autonomously exploit vulnerabilities at a higher level than previous

Researchers detailed a now-corrected prompt injection issue that allowed Apple’s on-device large language model to circumvent restrictions and execute attacker-controlled actions.

A new AI News report argues that modern AI systems create attack surfaces traditional security models were not built to handle, pushing security teams toward AI-specific controls and thinking.

OpenAI publishes its design principles for protecting AI agents against prompt injection and social engineering in agentic workflows.

Anthropic has raised alarms about industrial-scale distillation attacks against its Claude AI model, where competitors and third parties systematically extract Claude's capabilities to train cheaper rival systems. The revelation highlights a growing threat to AI companies' intellectual property and business models.