AI Engineer · October 6, 2026

Agent sandboxes: keep credentials outside the harness and scope tools by stage

video Agent Security AI Engineering

Agent sandboxes: keep credentials outside the harness and scope tools by stage video thumbnail

Why it matters

Jim Clark’s architecture separates research, fact-checking and publishing into stages with different permissions. Sandboxed agents use permitted tools through a gateway, while credentials remain outside the agent environment; signing authority belongs only to the stage that needs it. The design aims to reduce persistent secrets and limit what a compromised stage can do. A gateway still grants consequential capabilities, so credential separation must be paired with enforced tool and network policy. The talk presents an architectural pattern, not proof that any default installation provides these guarantees.

My takeaway: Give each stage only its required tools and egress destinations. Test that the research stage cannot publish or obtain signing material, that credentials do not enter model-visible files, and that temporary authority ends when the authorized operation completes.

Keep exploring

More curated notes connected through Agent Security and AI Engineering.

OWASP GenAI Security Project · guide

OWASP Top 10 for Agentic Applications for 2026

OWASP's community guide organizes agentic-system risk into ten categories, including goal hijacking, tool misuse, identity and privilege abuse, memory poisoning, insecure inter-agent communication, cascading failures, and rogue-agent behavior. It provides a shared taxonomy and mitigation starting point rather than a certification checklist or evidence that a deployed system is secure.

Microsoft Security Blog · guide

AI vulnerability research: measure reproducible findings and completed fixes

Microsoft’s FORGE account describes the work between a model’s vulnerability claim and a useful repair: reusable builds, duplicate removal, reachability checks, project-specific verification, reproducible triggers and regression tests. Structured rejection reasons help improve later searches. The useful operational measure is the flow of findings that survive verification and reach a fix, rather than the number of candidates generated. Reported successful-case costs exclude parts of screening, failed attempts and human work, so they are not the total cost of operating this pipeline.

OpenAI News · framework

Frontier training safety cases: connect evidence to enforced pause and rollback controls

OpenAI proposes training-run safety cases combining alignment evaluations, containment and monitoring with explicit operational ownership. Concrete measures include immutable transcripts, held-out incident tests, checks for evaluation gaming, response deadlines and fail-closed monitoring. Independent internal challenge, leadership vetoes and tracking downstream uses support stopping a run and reversing affected work. The article describes recommendations still being implemented, rather than audited proof that every safeguard already operates or that residual risk has been eliminated.