Vulnerabilities, prompt injection, model safeguards, and containment design.
Claroty introduced Claire, a CPS-native AI security agent. The launch shows why AI action in factories and hospitals has to be tied to asset data, approvals, and audit trails.
OpenAI Secure MCP Tunnel gives ChatGPT, Codex, Responses API, and AgentKit an outbound-only path to private MCP servers without public endpoints.
Anthropic published Claude Mythos Preview exploit evaluations and a CVD dashboard. V8 21/41 ACE and 1,596 disclosed flaws reset security triage expectations.
Anthropic detailed the isolation design behind Claude Code and Claude Cowork. The numbers turn agent security from approval prompts into sandbox, VM, and egress policy.
YouTube is moving AI-generated video labels onto the player surface and adding automatic detection for realistic AI media. Here is what changes for creators, viewers, and AI product teams.
Decepticon 1.1.3 shows that red-team agents are now competing on rules of engagement, sandboxing, graphs, release integrity, and auditability.
OpenAI outlined its 2026 election safeguards, combining AP vote counts, voting information, Codex Security, SynthID, usage policy, and political bias evaluations.
TELUS Digital tested 34 AI models with more than 620,000 adversarial attacks. The benchmark shows why enterprise AI safety is now an operating discipline.
TrapDoor combines malicious npm, PyPI, and Crates.io packages with poisoned AI coding instruction files.
Anthropic’s Claude containment writeup shows agent security shifting from prompt defenses and approval dialogs toward runtime isolation and blast-radius control.
OpenAI’s unit-distance counterexample shows that AI research automation depends less on answer generation than on proofs experts can inspect.
SLEIGHT-Bench uses 40 synthetic attacks to show how easily LLM monitors can miss risky behavior by coding agents.