Air Canada paid $812 for arguing its chatbot was 'a separate legal entity.' Two years later agents transact on live card networks — and California made the autonomy defense illegal. The honest answer: liability follows the evidence gap, and the evidence is your mandate and handoff logs.

CMS's position in one sentence: algorithms may inform care and coverage decisions but may not make them alone — and the enforcement mechanics assume you can prove which decisions your AI touched. The four instruments, mapped.

Both halves are true: Claude's enterprise controls are genuinely good (inference hooks, Compliance API, HIPAA path), and a sophisticated customer still presented a real Claude Code credential-exfiltration incident at Black Hat. The review covers both.

ChatGPT Work is not a bigger chatbot — it's an autonomous process acting under employee identity across your SaaS estate and the open web. The security review to read before saying yes: real credit for OpenAI's controls, five surviving risks, eight hardening steps.

A July 1 licensing change made this urgent: tenants without Agent 365-eligible licensing lost agent discovery in Defender. What Microsoft's agent stack genuinely does well, the four gaps, and why serious agent estates run both layers — like a CMDB and an EDR.

Every Microsoft-heavy security team asks this within twenty minutes. The honest answer: what Purview genuinely does well, the five places it structurally stops — non-Microsoft AI, runtime MCP connectors, endpoints, fragments, labels — and a decision table.

In July 2026, every one of the 79 enterprises Harmonic monitors had employees using China-based AI tools — M&A plans and source code included. The honest answer splits in two: the hosted service, and the open-weights models. Conflating them is the mistake.

The organizations getting burned aren't the ones that allowed AI — they're the ones that believed a ban worked. The behavioral data, the three postures honestly compared, and the six clauses of a policy that survives contact with reality.

The ChatGPT your employees use on personal accounts can never touch PHI. The enterprise tiers can — with a BAA and the safeguards most guides skip. Here is the tier-by-tier answer and the realistic path for healthcare organizations.

Every AI security incident that made news this summer had something in common, and it wasn't the browser. Our third product secures the AI blind spot — desktop apps, IDEs, local models, CLI agents, and MCP servers — with one policy that syncs across every surface.

Copilot doesn't break permissions — it surfaces every mistake you've made since 2014. Seven tools that find the oversharing first, described the way a buyer would want: what each does, where each stops, and which two you already own.

"Fewer breaches, probably" does not survive the budget meeting. This is the model that does: four value lines with named inputs, a worked example that clears platform cost with the soft line excluded, and the 40%-cancellation counterargument answered head-on.