What it evaluates
Each evaluation starts with the task and requested effect: what the agent can access, who authorizes changes, how limits are tested and what happened. The report identifies the reviewed scope and supporting evidence.
SecureStamp Agent Ethics Standard · v0
A public standard for delegating work to agents.
SAES describes the task, the authority granted, the safeguards applied and the behavior observed. It evaluates the delegation setup, not the moral character of a model.
What evidence means
SAES separates declared design, applied controls, demonstrated properties and observed behavior. Each conclusion shows the type and level of evidence behind it.
Why the safeguards matter
METR’s investigation of the Hugging Face incident documents persistence on difficult tasks, attempts outside assigned scope and gaps in observation. SAES turns those lessons into criteria for reviewing limits, oversight and traceability.
METRPrinciples
SAES applies seven principles: scope, legitimate authority, meaningful human oversight, proportionality, honesty and evidence, privacy, and moderation.
Evidence levels
N1 reviews the design; N2 tests safeguards in a controlled environment; N3 requires integration with the real destination.
Knockouts
Seven non-waivable requirements: D1-01, D1-02, D2-01, D2-02, D3-01, D4-01 and D5-01.
SAES profile
The profile groups 42 requirements across seven dimensions. Each evaluation identifies the controls that apply to the task and the evidence supporting the review.
Explore SAES requirements and specificationOpen changes (RFC)
Propose changes with their rationale, affected requirements, evidence and compatibility impact. Preserve disagreement and prior versions; generated support does not count as community consensus.
Open RFC discussionsChallenge a requirement or result
An appeal should identify the requirement or evaluation, the disputed claim and the evidence to review. An appeal does not change a decision automatically; the review and its outcome should remain traceable.
Open an issueAlternative evaluators
The RFC process lets teams propose alternative evaluators. Proposals document methods, evidence provenance and reproducibility, conflicts of interest and limits.
Evaluation scenarios
SAES covers risks such as cross-run contamination, exposed credentials and derived execution. Each scenario examines the boundary, the correction and whether a legitimate task can continue.
Two users, the same evidence
Development uses SAES to deliver a change; security uses it to decide what it can authorize. Both see the same facts and answer different questions.
What runs after the change
A workflow, dependency, script or agent instruction can set up a later run with different permissions. SAES shows what runs next, when and with which declared permissions; routes outside observation are clearly marked.
Evidence you can verify without SecureStamp
Each review exports as a bundle that verifies offline, including with OpenSSL. The bundle lets teams check file integrity, signatures and the scope of observation.
Public method, consistent criteria
SecureStamp maintains SAES and offers evaluations. Each result follows the same requirements and evidence regardless of plan; the report shows what supports the conclusion.