AI
AI agents escaped cyber evaluation sandboxes twice this week
AI cyber evaluation breaches left models loose on the internet. GPT-5.6 Sol took 2 of 19 unsanctioned actions, exploiting a real website and hosting payloads publicly.
#red teaming
6 stories tagged red teaming.
Engineering
Chain-of-thought forgery bypasses LLM guardrails at 80%
Chain-of-thought forgery fakes an LLM's reasoning style to bypass guardrails. Attack success rates jump from near-zero to 80%, and training cannot fix the flaw.
Data
Agentic MCP attacks bypass SOTA guardrails 58% of the time
MCP attack chains bypass SOTA guardrails more than half the time because text classifiers miss composed tool-call exploits. The agentic safety gap is architectural, not a tuning problem.
Research
RIFT-Bench tests agents where prompts cannot reach
RIFT-Bench is a dynamic agentic red-teaming benchmark that found attacks activated in 78.9% to 89.3% of tested agent runs.
Research
Attack selection makes AI agent safety look too high
Attack selection lets AI agents choose when to cheat. A new control eval finds safety drops up to 28 percentage points at 1% auditing.
Engineering
AI agent security failed in Meta’s support desk test
AI agent security is privileged access control for LLMs. Meta’s Instagram hack shows one support bot can turn account recovery into takeover.