Open source· forensic evidence for AI agents
Mnestiq records what your agents saw, decided and did as signed, tamper-evident evidence. When something goes wrong, you can name the run, the step, the input that drove it and who approved it, and prove the record wasn't altered afterwards.
pip install mnestiq
- Firewall10.0.4.17:53120 → 198.51.100.9:25 · 48 KB · no agent identity
- Mnestiqagent support-bot · run f48cccfa · step 6 · sandbox container:a91f02 · operator acme-helpdesk
- Causeuntrusted input at step 4 (web, via fetch_page) → model requested send_email → approved by policy:email-auto-approve
- Evidencesealed by checkpoint · signer ed25519:6bec5262… (pinned) · corroborated by vpc-flow.log, conn.log
The gap
Your logs see connections. Your traces see prompts. Nothing connects them. #
A firewall alert: a host that runs AI agents sent 48 KB to an outside mail server on port 25. The responder now has three questions, and today's tools answer none of them.
links them
Which agent did this?
Firewall and flow logs show an IP and a port. They don't know about agents, runs or steps.
Why did it do it?
Was the instruction from the user, or from a web page, a document or another agent? That has to be captured when it happens, not reconstructed later.
Can we prove it?
Ordinary logs can be edited by anyone with write access. Evidence for a post-mortem, an insurer or a regulator needs to show it wasn't.
How it works
Record, verify, investigate. #
A flight recorder in the agent
Records agents built on Anthropic, OpenAI, Azure OpenAI and OpenAI-compatible models in a few lines. Built to fail safely: if recording fails, the agent keeps running.
- Every piece of model input tagged by source: user, web, tool output, retrieved document, another agent
- Tool calls, approvals and effective permissions
- Every outbound connection, with run identity headers and the local address and port
Marked sources are treated as untrusted by the detection rules.
import anthropic
from mnestiq import FileSink, Recorder
from mnestiq.egress import instrument_egress
from mnestiq.integrations.anthropic import instrument_anthropic
from mnestiq.keys import load_private_key
recorder = Recorder(FileSink("evidence.jsonl"), agent_id="support-bot",
signing_key=load_private_key("keys/signing.key"))
client = instrument_anthropic(anthropic.Anthropic(), recorder)
instrument_egress() # stamp and record outbound connections during runs
with recorder.run():
... # your agent, as it is today
recorder.close() # signs the last checkpoint: verify needs it
Tamper-evident by construction
Records are hash-chained and sealed by Ed25519-signed Merkle checkpoints.
- Edits, deletions, insertions and reordering fail verification
- Pin the recorder's key, and a re-signed forgery is rejected
- Redact prompts and customer data before sharing, and the proof still holds
record hash of the record before
record hash of the record before
record hash of the record before
Signed checkpoint Merkle root of the records it covers · Ed25519 signature
Change, remove, add or reorder a record and the hashes stop matching what was signed.
From a log line to a cause
Join agent evidence with the infrastructure logs you already collect.
- VPC flow logs, Zeek and proxy logs linked to the exact agent step, with a confidence score
- The untrusted input that preceded an action, and who approved it
- Traffic from agent hosts that no recorded step explains
When someone edits the record
Built for the person holding the incident. #
What the responder gets
The run, the step, the input and the approval.
The same evidence file answers the three questions: which agent, why, and whether the record can be shown to be unaltered. Anyone can check it with the open-source verifier, without Mnestiq.
Two products
Open evidence. Professional detection and investigation. #
Mnestiq recorder
The evidence format, recorder, verifier and a local dashboard. Free and open source, so you can audit it yourself.
- An open spec and JSON Schema, aligned with OpenTelemetry GenAI conventions
- Python SDK for Anthropic, OpenAI, Azure OpenAI and OpenAI-compatible models
- Signing keys held outside the agent, in Azure Key Vault or a dedicated signing service
- Independent RFC 3161 timestamps and signed heartbeats
- Crash-safe, size-capped, fails open by default
- Dashboard runs on your machine only
Mnestiq Workbench
For security teams that need to see what their agents do as they run, and to investigate and prove what happened when something goes wrong.
- Live detection: alerts in Microsoft Sentinel within minutes, every agent's activity, and a dashboard in the tools your SOC already uses
- Agent evidence joined with flow, DNS, proxy, firewall and cloud logs, with a stated reason for every link
- Microsoft Sentinel both ways: logs come into the case, triage decisions go back onto the incident
- An independent Evidence Vault that catches evidence cut back or rewritten
- Signed reports, and a claim pack written for insurers that anyone can check
Security model
What it proves, and what it doesn't. #
Evidence is only useful if its limits are known. We state them plainly in the open spec and in a published threat model, where every attack considered is backed by an automated test.
Evidence Vaultdoes this for you.
Verify, don't trust.
You should not have to take a security vendor's word for it. The recorder and its verifier are open source, every claim in the threat model is backed by an automated test, and any evidence file can be verified independently, without Mnestiq. Prompts and data stay on your own systems: only signed checkpoint fingerprints leave, and only to destinations you configure.
Independent review is welcome. The break-it challenge provides a signed, timestamped incident record: alter it without the verifier noticing, and tell us how.
Design partners
Running agents with real permissions in production? Let's talk. #
We're looking for a small number of security and platform teams to shape the workbench around real incidents.