Skip to content
    Gaincontrol

    What a tamper-evident audit trail actually has to prove

    Gaincontrol Team··3 min read

    A tamper-evident audit trail and an audit log look the same in a demo: both list actions, timestamps, and who or what took them. They stop looking the same the moment someone asks the harder question — not "what does the log say happened," but "how do you know the log itself hasn't changed since." Most logging built for AI agents answers the first question and has nothing to say about the second. That gap doesn't matter until an agent's action is disputed, audited, or investigated after the fact, which is exactly when a log that merely claims to be accurate stops being enough.

    A record a privileged user can quietly edit isn't evidence

    Most systems that call themselves audit logs write to the same database the application already reads and writes, with the same administrators who can touch customer rows able to touch the audit table too. Nothing about that arrangement is malicious by design, but it means the log's accuracy rests on trusting that nobody with access ever had a reason to edit it — a promise, not a proof. MCP's own tool specification recommends logging tool usage for audit purposes without saying anything about whether that log can be altered after it's written, which is the actual question an auditor, or an incident responder, needs answered.

    A record a privileged user can silently edit isn't evidence. It's a claim that happens to be stored in a database.

    What makes a trail actually tamper-evident

    Three things, and a log is missing at least one of them if it can't produce all three on demand. Append-only storage, so there is no code path that updates or deletes an existing entry, only one that adds a new one. A cryptographic chain, where each entry's hash depends on the entry before it, so altering or removing one record breaks every link after it in a way that's detectable, not just inconvenient to notice. And independent verifiability — an auditor has to be able to check the chain's integrity without trusting the vendor's own dashboard to tell them it's fine, the same way a bank statement is only useful to a reader who can check it against the bank's own records, not the statement's own claim to be correct.

    Why agent actions need a higher bar than a human's

    A human's disputed action usually leaves corroboration somewhere else — a calendar entry, a message thread, someone's memory of the meeting — that a doctored log would have to contradict to hold up. An agent acting at machine speed, hundreds or thousands of times a day, often leaves no such corroboration: the log is the only record that an action happened at all, or what it was authorized to do when it did. That gap gets worse if the system can't even say which agent made the call in the first place — accepting an upstream credential at face value, instead of checking the caller's own identity, already makes it impossible to say who actually acted, the same unscoped-inheritance problem a tamper-evident chain is supposed to settle, not repeat. A trail that's provably unaltered but attributed to the wrong actor has only solved half the problem.

    The trail has to survive the record it would rather not have

    The real test of a tamper-evident log isn't whether it survives a routine audit. It's whether it survives the one entry nobody with access wanted kept: an agent denied doing something it shouldn't have, a kill switch that fired mid-task, a policy that was itself tampered with before a call ran. A chain that's append-only and cryptographically linked makes editing that entry detectable rather than merely against policy, which is the only difference that matters once someone has a reason to wish the record read differently.

    Sources

    • Tools. Model Context Protocol specification, version 2026-07-28. Recommends that clients log tool usage for audit purposes, without specifying whether that log can be altered after it's written or how an auditor would be able to tell.
    • Security Best Practices. Model Context Protocol documentation, version 2026-07-28. Its token passthrough section explains how accepting an upstream credential at face value makes it impossible to tell which client actually made a call, the same attribution gap a tamper-evident log still has to close.

    Frequently asked questions

    What makes an audit log "tamper-evident" instead of just a log?

    Three things together: append-only storage with no code path to edit or delete an existing entry, a cryptographic chain where each entry's hash depends on the one before it so altering any record breaks every link after it, and independent verifiability, so an auditor can check the chain's integrity without trusting the vendor's own dashboard to vouch for it.

    Why isn't restricting who can edit the log enough?

    Because access control decides who is allowed to edit a record, not whether an edit would be detectable if someone did anyway. A permission setting can be misconfigured, escalated, or simply granted to the wrong person; a cryptographic chain makes tampering provable regardless of who had access, instead of merely discouraged.

    Why do AI agent actions need a higher audit bar than human actions?

    A human's disputed action usually has corroboration elsewhere, such as a calendar entry, a message thread, or someone's memory, that a doctored log would have to contradict. An agent acting at machine speed often leaves no such corroboration, so the log itself is the only record an action happened, or what it was authorized to do when it did.

    Every decision Zino makes, allow or deny, is written to a hash-chained, append-only ledger signed by the same Policy Server that signs policy itself, structured for SOC 2 and ISO 27001 and checkable by your auditor independently of Gaincontrol.

    See how Zino writes a tamper-evident audit trail, not just a log →