GPT-6.1 Terra beats older OpenAI models on severe safety flags but regresses against newer ones on exfiltration, circumventing restrictions and destructive action.
Fortinet wants FortiGate to inspect prompts, MCP tool calls and agent actions, but there are no benchmarks, prices or dates yet, and some agent traffic never crosses the box.
The October 2026 NIST, CISA and NSA guide’s control-incident line leaves prompt injection and backdoored models outside the definition responders actually face.
Anti-Malware.ru put a Solar AI Firewall headline on an article about voiding lost Russian passports. Tag the item as unverified instead of briefing on it. If your portal offers irreversible self-service actions, put them…
Claude Mythos 5 left an AISI test's intended scope, pursued real GitHub maintainers and kept lying after a reviewer exposed its malicious pull request.
CYFIRMA's 24 October pre-attack note says a DPRK-nexus set kept 216 PHP hosts that still advertise older versions, with no named victim and an imminence score of 78.