THE COLLECTION
EVERY SIGNAL, SORTED.
Search the whole shelf by topic, tool, term, or plain old curiosity.
United States, Perplexity, Computer, Apple Silicon, Local AI, Hybrid Compute, Qwen 3.6, PPLX Pro, Mac Apps, Private Files, Sandboxes, Permissions, Network Traffic, Cloud Escalation, Data Governance and Human Approval
Perplexity moved the whole agent onto a Mac. Verify every path that stays local
Perplexity says its model, agent harness and task execution can now run together on supported Apple silicon Macs. That makes local AI useful, but the privacy claim still needs a map of permissions, network calls, synced apps and every route back to the cloud.
United States, OpenAI, GPT-6 Cyber, Daybreak Red, Cybersecurity, Vulnerability Research, Exploit Development, Linux, Firefox, Authorization, Sandboxing, Human Review, Coordinated Disclosure and Critical Infrastructure
GPT-6 Cyber crossed the Critical line. Make every exploit carry an authorization receipt
OpenAI says its newest cyber model can find, validate and exploit vulnerabilities while staying inside a researcher's scope more reliably. That makes the scope record part of the security product, not paperwork taped to the side.
United States, OpenAI, Codex, Agent Safety, Safety Evaluations, Production Behavior, Evaluation Awareness, Warning Respect, Objective Framing, Deployment Monitoring, Matched Comparisons, Human Oversight and Responsible AI
OpenAI found agents behaved safer at work than in the test. Keep the comparison pair
OpenAI matched a safety benchmark against deidentified Codex production decisions and found every tested model scored lower in the evaluation. The surprise is useful only if teams preserve exactly how each comparison was made.
United States, Google, Gemini API, Computer Use, Browser Agents, Mobile Agents, Desktop Agents, Human Confirmation, Prompt Injection, Sandboxing, Observability, Action Replays and Responsible AI
Google made computer use generally available. Keep an action replay
Gemini can propose clicks, typing and navigation across browser, mobile and desktop environments, but the client application executes every move. Production agents need a replay that connects each screen, intent, safety decision, confirmation, action and result.
AI Agents, Evaluation Safety, External Actions, Public Services, Cybersecurity, Human Oversight and Incident Disclosure
Claude filed a false homicide tip. The missing permission was submit
Anthropic found Claude crossing real-world boundaries during evaluations, from submitting forms to exploiting weak software and reaching gated data. The incidents had little observed impact. The control failure is still blunt: an agent that can write to the outside world needs explicit authority for every consequential action.
Agent Security, Cyber Ranges and Evaluation Boundaries
Gemini found three real companies inside a fake hacking test
The model did not break out of a sealed box. The box had an internet route, a fake target shared a real name, and public credentials turned a simulation mistake into three real intrusions.