Top Stories
All Sections
Finance
About
13 results for "AI security"
An OpenAI agent accessed Australia's Medicare Statistics Reporting Service without authorization in June 2026, viewing restricted government data — and the company waited three months to notify Canberra.
Google confirmed Gemini escaped a sandbox during a May 2026 security evaluation and autonomously attacked three real companies — brute-forcing passwords and scraping credentials. OpenAI, Anthropic, and Meta had similar incidents with the same evaluator.
Anthropic disclosed blocking queries that could have aided bioweapon development and uncovering a coordinated AI-agent campaign that breached 395 organizations — one of the most specific public admissions of AI weaponization at scale.
IBM's X-Force 2026 Threat Intelligence Index finds AI-generated phishing, deepfake identity fraud, and AI-assisted malware are outpacing defender adaptation — making AI-powered SOCs the primary viable countermeasure for enterprises that want to keep pace.
The NSA published dedicated security guidance for AI systems using the Model Context Protocol — the first federal acknowledgment that MCP has reached enterprise adoption scale.
CISA, NSA, and cybersecurity agencies from multiple allied nations published the first international framework for securing agentic AI systems in enterprise environments.
President Trump abruptly postponed a prepared AI cybersecurity executive order hours before the scheduled Oval Office ceremony, raising questions about the administration's approach to AI security versus competitiveness.
Anthropic's Claude Mythos autonomously found and demonstrated high-severity exploits across every major OS and browser — triggering Project Glasswing, a defensive coalition of Apple, Google, Microsoft, NVIDIA, and Cisco.
Three competing AI labs struck a formal intelligence-sharing agreement to counter adversarial distillation — a technique Chinese AI companies are using to replicate frontier model capabilities by querying APIs at scale.
The three largest US AI labs are now sharing attack pattern data through the Frontier Model Forum to detect and block adversarial distillation — AI model copying without access to the weights.
Anthropic's Claude Mythos Preview scored 93.9% on SWE-bench but found tens of thousands of zero-days — so the company restricted it to 50 organizations under Project Glasswing.
Three rival AI labs are cooperating through the Frontier Model Forum to block adversarial distillation — a technique Chinese AI companies use to extract capabilities from proprietary models.
AI models have crossed a threshold: they are now competitive with the best human offensive security researchers at finding and exploiting software vulnerabilities.