Practical guide. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Developers and teams putting AI into practical use
⚡ TL;DR — the 30-second version
- Provider safeguards do not replace least privilege.
- Treat retrieved instructions as untrusted content.
- Gate consequential actions and retain redacted traces.
Read the model's system card, but do not outsource security architecture to a benchmark. The Opus 5.5 report discusses stronger alignment evaluations alongside limitations and domain-specific safeguards. No such report proves that a particular deployed agent cannot leak data or make unauthorized changes.
Use least-privilege tools, separate credentials by environment and redact secrets in traces. A research agent should not possess production deployment credentials. Retrieved documents and websites are data, not higher-priority instructions.
“A safer model is one layer. Least privilege and human authority are properties of the application you build around it.”
Require approval for consequential actions and make the proposed change readable. Record what the agent attempted, what policy allowed and what actually executed. Budgets, cancellation, timeouts and an incident runbook should exist before unattended runs.
Retest after a model or harness change. More capable reasoning can change both task completion and the path an agent takes to reach its answer. Include adversarial input and permission boundaries in the regression suite.
Field guide / 01
Build a permission matrix
List every tool and resource, who can request it, what data it receives and whether it can write. Make the most dangerous path explicit: a low-trust document influencing an agent that holds a high-trust credential.
Remove unnecessary credentials and constrain the execution environment before refining prompts. A policy that prevents access is more reliable than a reminder asking the model not to use it.
Field guide / 02
Rehearse a contained incident
Use synthetic secrets and a disposable environment to test malicious retrieved instructions, cross-user access and unexpected network requests. Confirm the enforcement layer denies the operation, records the attempt and alerts the right person.
Define cancellation and reconciliation for in-flight actions. Stopping generation does not reverse a database write or recall a sent message. Your incident response needs the actual execution state.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Questions that come up
Can a prompt alone prevent tool misuse?
No. Use prompts as guidance, with authorization, isolation, scoped credentials and validation enforced by the application.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01Claude Opus 5.5 system cardAnthropic↗
- 02Introducing Claude Opus 5.5Anthropic↗
- 03OpenAI API changelogOpenAI↗
- 04MCP — security best practicesModel Context Protocol↗
- 05Codex — agent approvals and securityOpenAI↗