Agent security evidence
Your deal is waiting on a security review. Hand them proof.
kolm audits your AI application - the agent and every tool, identity and data flow around it - from the logs you already have, and signs the findings with Ed25519. Your buyer verifies the report in their browser, offline. kolm is never in the trust path.
The reviewer wants proof. You have a questionnaire.
Three findings from the signed sample. Every claim below sits inside the signature, so a buyer checks the evidence, not our summary.
More access than the job needs
The first thing a reviewer flags, and the hardest to see from inside.
A log nobody can trust
If the trail can be edited after the fact, it is not evidence.
Say-so does not scale
Self-attested answers shift the burden of proof to the buyer.
Audit. Sign. They verify.
One lifecycle, on demand, in minutes. Reproducible: same logs, same result.
Run it on your logs
Import a log or attach the sidecar proxy. Permissions, audit trail, egress and injection resistance are read from what the agent did.
We seal the findings
The report is canonicalized and signed with Ed25519. The signature covers the exact bytes: inflate a score and the seal breaks. Every issuance enters an append-only log.
Your buyer checks it
Offline, in their browser, against the public key inside the report. The check needs only the report file; kolm never sees it happen.
Eight controls read from your own logs · Six frameworks in the crosswalk · mapped to the controls your buyer already cites · the full lifecycle
Read your logs
Import a log file or attach the sidecar proxy. Nothing leaves your runtime to start.
12400 calls readThe controls run
The controls are graded against scope. Each finding cites a control and a framework clause.
13 findings, 7 highThe seal goes on
The findings are canonicalized and signed with Ed25519. Change one byte and the seal breaks.
Ed25519 one signed objectThey check it offline
Your reviewer opens the report in their own browser and checks the signature against the key inside it.
VALID no server of oursThey do not trust us. They check the signature.
Two checks run in the buyer's own browser against the key inside the report. kolm is never in the verification path and grades nothing it sells.
Tier 1 · signature
Edit one field, the seal breaks
The signature covers the canonical bytes. Change a finding or inflate the score and the match fails, in front of the reviewer.
Tier 2 · issuer
A rogue key signs, but does not match
The signing key is pinned to the keyring your buyer expects. A forged key clears tier 1 and fails tier 2.
Offline · no server
Nothing for us to fake
Verification runs on WebCrypto in the buyer's browser. No account, no upload, no server of ours in the path.
Scope is contractual. Permission posture, redaction and audit-trail integrity are assessed. Injection is tested and reported, not warranted.
Issuer key fa562154f99c95f4... · append-only transparency log · inspect the verifier source
What a Signed Readiness Report actually contains.
Not a PDF and not a slide. One canonical (one fixed, byte-for-byte ordering)A canonical object is serialized one fixed way - keys sorted, no stray whitespace - so the same report always produces the same exact bytes, and the signature can cover those bytes. JSON object that carries its own scope, its own findings, and the key needed to check it. Here is every part, and what each part is for.
Helpwise support & billing agents, 12,400 calls.grants 10 tools, uses 4 on one shared key.The free Scan signs the same object with a watermark inside the signed bytes; the $750 Signed Readiness Report re-signs the same audit with the watermark removed. Walk the object field by field
Start free. $750 when the deal needs a signature.
Flat fees, self-serve, from a free scan to a guided fleet audit. Open the platform comparison to see how signed evidence connects to the wider Kolm control loop, or estimate what a stalled review costs.
Unstick the deal.
Run the scan tonight. Hand the reviewer signed evidence tomorrow.
Caveats: Scope is contractual. Permission posture, redaction and audit-trail integrity are assessed. Injection is tested and reported, not warranted.