A few weeks ago, OpenAI's GPT-5.6 Sol escaped a sandbox during a safety evaluation and reached Hugging Face's production systems. We wrote up what it says about agent authorization, and covered research showing safety-aligned models refuse defensive security requests 2.72 times more often than neutral ones. The model is not a reliable place to put the access decision.

For security leaders, we looked at what a single policy decision point does for incident response, where containment is a policy change instead of a code deploy and the evidence trail is already there. We also covered DORA Article 9(4)(c) and AI agents, and published our security and trust page for teams running us through vendor review.

For builders, we shipped agent skills for writing policies in Codex CLI, Cursor and VS Code. Plain English requirements in, a bundle validated against the real Cerbos compiler out, human review still required before it ships. Our base policy authoring guide covers the full workflow and links out to the setup for every coding assistant we support a skill for.

On MCP, a gateway gives you routing, not per-agent tool authorization, so we went through the standards that cover that.

We also refreshed RBAC vs ABAC, and wrote up guides on authorization in microservices, as well as mapping business requirements to authorization policy in the gaming space.

To close, we brought audit log exports to Cerbos Hub, updated the Playground to show compiler errors inline, and shipped two Cerbos PDP releases: v0.54.0 with faster evaluations, lower memory use and new store diagnostics, and v0.55.0 with an opt-in strict evaluation mode, support for multiple JWTs in one request, and faster conditions.

🤝 We launched a referral program

Thank you to everyone who has sent teams our way, it is how some of our best conversations started. Know another team wrestling with authorization? Submit a referral, and if that company becomes a Cerbos customer, that is $1,000 to you. It takes two minutes. Refer a team here.

Product Updates

Cerbos Hub

The latest Cerbos Hub update gives teams audit evidence on demand, and a Playground that shows exactly what is wrong with a policy before it ships.

Audit log exports are now available to workspace and organization owners, straight from the Audit logs page, carrying whatever time range and filters are already applied.

In the Playground, compiler errors now show inline at the offending line, undefined variables, ambiguous derived roles and scope issues get named, and fixtures are validated against their attached schemas.

Learn about Cerbos Hub

Get started with Cerbos Hub

Cerbos PDP

The latest Cerbos PDP releases give organizations stricter failure behavior when a policy expression breaks, and faster evaluations.

Cerbos PDP v0.55.0 introduces an opt-in strict evaluation mode: a condition or variable that raises a runtime error is treated as unsatisfied today, which skips a DENY rule and continues evaluation without it, whereas strict evaluation denies the affected action instead. Conditions are also faster, with constant sub-expressions folded at compile time rather than evaluated on every request, and auxData now accepts a set of named JWTs, each with its own keyset. The CEL upgrade changes some PlanResources output, so deployments that translate query plans into database queries should compare plans before and after upgrading.

v0.54.0 brought lower memory use, GOMEMLIMIT set automatically from cgroup limits, and a cerbos compile-store command for diagnosing blob and database stores.

You can find the full release notes here: v0.54.0, v0.55.0
Cerbos PDP is open source, check out our GitHub

Helpful Content

IAM security checklist for 2026

IAM gaps remain one of the most common root causes of breaches, and most teams don't have a clear picture of where they stand. 

This checklist is a practical, priority-ranked assessment of the IAM controls that matter most in 2026, built from what we've learned working with enterprise security teams, industry experts, and leading identity events including Gartner IAM, Identiverse, and EIC.

Upcoming Events

Visit our booth at: 

• Identity Week America, Washington DC, September 2-3. Stop by the Cerbos booth 720 for a proper chat about all things authorization. We also have a few VIP guest passes to give away. If you'd like one for yourself or a colleague, email aram@cerbos.dev and he'll let you know if any are still available.

Meet our team members and check out their talks to get valuable insights:

• MCP Dev Summit, Seoul, August 13. Stop by Aram Andreasyan’s talk “Authorization in MCP systems: Getting it right from the start”.
• TailscaleUp, San Francisco, August 26. Check out Alex Olivier’s presentation “Every tool call is a trust boundary: Authorization for AI agents” which includes  a live demo built on Aperture and Cerbos.
• ISC2 Security Congress, Gaylord Rockies, October 24-28. Check out Alex's talk “Identity as the control plane for software supply chain security” on October 28, co-presented with Vatsal Gupta of Apple.