ProvenantAI Get in touch
← Standards register Assessment

CSA Agentic Trust Framework

Agent trust framework · 25 requirements

Instead of learning your baseline from traffic and writing it down, Provenant defines and enforces it in advance for you.

22

Enforced

Provenant does it. You can show it before your agent runs.

2

Contained

Where the standard asks you to prevent something the model might do, Provenant limits how far it gets.

1

Evidenced

Where the standard asks you for a plan, a policy or a test, Provenant hands you what it needs.

RefRequirementHowMechanism
I-1Unique identifier per instanceEnforcedEvery session is identified and signed
I-2Bound to cryptographic credentialsEnforcedEvery session is cryptographically bound
I-3Ownership and responsibility chainEnforcedResponsibility stays traceable even when agents call agents
I-4Documented use and scopeEnforcedWhat you set up is the scope
I-5Machine-readable capabilitiesEnforcedYour agent's options change with where it is in the work
B-1Machine-parseable action logEnforcedRecords are written and signed as they happen
B-2Tied to identity and sessionEnforcedEvery action ties back to a session and an identity
B-3Baseline of normal operationEnforcedProvenant defines and enforces your baseline for you
B-4Deviation detectionContainedThere is nothing to deviate to
B-5Retrieve decision rationaleEnforcedYou can see why any decision was allowed
D-1Input structure conformanceEnforcedInputs are checked before they reach the tool
D-2Prompt-injection resistanceContainedAn injection can't reach what your agent was never given
D-3Sensitive-data controlEnforcedYou choose what the agent sees, field by field
D-4Output structure conformanceEnforcedOutputs are checked, and the record proves it
D-5Data provenanceEnforcedEvery value traces back to where it came from
S-1Permitted resources enumeratedEnforcedAnything your agent can reach is listed up front
S-2Permitted actions enumeratedEnforcedListed for every step of the work
S-3Maximum operations per periodEnforcedYou can cap how much happens in a session
S-4Maximum impact per actionEnforcedYou can cap how much any single action does
S-5Cumulative and cascade limitsEnforcedRunaway loops and cascades are capped
R-1Automatic halt on repeated failuresEnforcedRepeated failures stop the agent automatically
R-2Immediate manual terminationEnforcedYou can stop an agent immediately
R-3Invalidate all sessionsEnforcedYou can invalidate every session at once
R-4Undo actionsEvidencedThe record gives you what you need to reverse things
R-5Fallback to lower autonomyEnforcedYou can drop an agent to a lower level of autonomy