Early access · Mnestiq Workbench, detection and investigation for AI agents

From an alert about an AI agent to a case you can hand over.

The Workbench watches the signed evidence your agents record as they run, raises alerts in Microsoft Sentinel within minutes, and joins the evidence with the network and cloud logs you already collect. It names the agent step behind every connection, explains why it happened, and produces evidence your SOC, your lawyers and your insurer can check for themselves.

Windows installer, Linux package or container · runs on your own machines · nothing leaves unless you send it
mnestiq workbench · one lead, explained
Mnestiq Workbench: a lead showing, in plain language, how planted web content led the agent to email the customer export
One lead, explained: the planted instruction quoted from the page the agent read, the email it led to, the policy that approved it without a person, the network records that confirm the connection, and what to do about it.
The evidence trail

One evidence trail, from the agent to the claim.

  1. Recorder

    Open source, inside the agent. Records every model call, tool call, approval and connection, signed with a key the agent can't take: in Azure Key Vault or a separate signing service.

  2. Evidence Vault

    Keeps an independent copy of each signed checkpoint in an append-only log, so the evidence can't be quietly cut back or rewritten.

  3. Workbench

    Joins the evidence with flow, DNS, proxy, firewall and cloud logs, and raises alerts as leads appear.

  4. Microsoft Sentinel

    Alerts arrive as incidents. Logs come back into the case. Decisions go back onto the incident.

  5. Reports and claim pack

    Signed, redacted, and checkable by anyone, with or without Mnestiq.

The case

Every connection, traced to the step that made it.

A case holds the originals, byte for byte, and everything the Workbench concluded from them. Analysts work in a local web app or on the command line. Both run the same code and leave the same trail.

  • Every link, with a stated reasonEach network record is linked to the agent step that made it, with a confidence score, a stated reason and any alternatives.
  • Clocks that disagreeA sensor whose clock runs a few seconds out doesn't break the story. The Workbench allows for the difference between each log's clock and the agent's.
  • What the agent didn't recordTraffic from agent hosts that no step explains, possible DNS exfiltration, and agent credentials used from elsewhere, each with what was running at the time.
  • Triage that leaves a recordEvery lead, most severe first, with what to do about it. Decide with a click or a key, one at a time or several together, with a reason and a name. An allowlist for known partners, notes, search, and the incident's discovery and containment times.
  • What it flagsA destination or account from untrusted content used in a tool call, actions approved by a policy instead of a person, connections after untrusted input, sensitive actions such as deleting, cancelling, sharing or changing access after untrusted input, and data leaked through an image or link in the agent's answer.
  • Set up in the browserA Set up page takes you through the licence, the recorder's keys and heartbeats, the Evidence Vault, Sentinel, alert destinations and the background service, and shows which steps are done. Secrets typed there are never shown again and never stored in the case.
  • Hand it overA triage bundle puts the slice of the case around one lead into one signed zip, readable in half an hour. A Timesketch export puts the agent's steps into the timeline your IR team already builds.
case overview
Workbench case overview: the leads that need a decision, alerts, evidence trust with the checks behind it, Evidence Vault receipts, incident times and recent activity
timeline · agent steps and network records in one sequence
Workbench timeline: agent steps and VPC flow, Zeek and proxy records in one sequence, each network record linked to its agent step, leads marked
Agent evidenceThe open-source Mnestiq recorder, verified against the agent's pinned key, collected once or followed live
NetworkAWS VPC Flow Logs, Zeek conn, dns, ssl and http logs, Envoy-style proxy logs, DNS query logs, CSV firewall exports
CloudAWS CloudTrail, Azure Activity, Google Cloud audit logs
From SentinelDNS, firewall (CEF), Azure Activity and AWS CloudTrail tables, or your own KQL, pulled straight into a case
Evidence trust

How far the evidence can be relied on, check by check.

For each agent's evidence, the Workbench gives a verdict (holds, needs review, does not hold) and every check behind it, each with how it was checked. The claims report and the claim summary carry the same checks, so a lawyer or a claims handler reads exactly what the analyst saw.

evidence trust · the verdict and the checks behind it
Evidence trust: the evidence holds, with each check listed: records unchanged, signed by a pinned key, signed coverage, not cut back, time proven, recorder running, held by the Evidence Vault
Signed outside the agentCheckpoints are signed by a key the agent's host can't take: a P-256 key in Azure Key Vault, where only a digest leaves the machine, or a signing service under its own account that refuses backdated or out-of-order checkpoints. A key change is handed over and signed by the previous key.
Time from outsideRFC 3161 timestamps from DigiCert, with Sectigo as fallback, prove the evidence existed by a given time, whatever the agent's clock said.
Silence showsThe recorder signs a heartbeat when idle, so a stopped recorder shows as a gap instead of looking like a quiet agent. A recorder that sends no heartbeats at all is marked for review.
Nothing cut from the endThe Evidence Vault's signed receipts show the evidence still reaches the latest checkpoint it received.
Tampering is a leadEvidence changed after collection, or refused by the Vault, becomes a lead with a decision, and a high-severity incident in Sentinel.
Microsoft Sentinel

Alerts go into Sentinel. Logs and decisions come back.

Sentinel sees every connection but not inside the agent. The Workbench sees inside the agent. Connected both ways, the SOC gets the why without leaving Sentinel, and the investigation gets Sentinel's logs without anyone exporting files.

Each lead becomes an incident

Alerts go to a dedicated table through the Logs Ingestion API, and a near-real-time rule turns them into incidents within about two to three minutes, grouped per case and tagged with MITRE ATT&CK techniques.

Every alert carries the plain-language story of what happened, and the command to open it in the Workbench.

Sentinel's logs in the case

Pull DNS, firewall, Azure Activity or AWS CloudTrail rows for a time window, or any KQL query, straight into an investigation.

Each pull is kept as an original, and the query and time range go into the chain of custody, so it can be repeated exactly.

Triage lands on the incident

Every decision made in the Workbench goes onto the Sentinel incident as a comment: who, what, why, and the custody record behind it.

An incident closes itself only when every lead in it is cleared. Confirmed incidents, and anything involving tampered evidence, are left for a person to close.

Your AI agents, in Sentinel

The "Mnestiq: AI agents" workbook shows alerts over time and what they are about, every agent with what it did (counts only, never content), each watcher and its heartbeats, problems with the evidence, and the incidents raised. A watcher that goes silent raises an incident of its own.

Sentinel setupAzure CLI
# One template deploys the tables, the ingestion rule, the analytics rules, the workbook and the app's permissions
az deployment group create --resource-group <rg> --template-file deploy/sentinel/mnestiq-sentinel.json \
    --parameters workspaceName=<workspace> appObjectId=<app>
mnestiq-wb watch ./fleet /var/log/agent-evidence --to sentinel:sentinel.json --heartbeat-minutes 5

No command line needed: the Set up page has a Deploy to Azure button that opens the same template in the Azure portal, filled in. The app's secret never goes into the case or its settings files. It comes from the environment, or from an owner-only file that only the Workbench service reads, and the Workbench refuses to send it anywhere but Microsoft's sign-in and ingestion endpoints.

Before and after

What your logs say today, and what they mean.

The same four connections, as the network logs record them and as the Workbench explains them. One was normal work, one fetched the planted instruction, one sent the customer export, and one was a 1.8 MB upload the agent's own record never mentioned.

before and after · four connections
Before and after page: each connection from the network logs next to the agent, step and reason the Workbench adds
Evidence Vault

A copy kept away from the agents' machines.

Rolled backThe evidence file is cut back to before the incident. The Vault already holds a later checkpoint, so the cut shows.
RewrittenRecords are changed and re-signed. The checkpoint no longer matches the one the Vault received.
Signer swappedThe evidence is signed with a different key. The Vault raises it the first time it sees the new signer.
ReceiptsFor each checkpoint, the Vault returns a signed receipt with the time it was received. The case keeps every receipt.
  • Why it mattersSigned evidence proves records weren't edited. On its own it can't prove nothing was removed from the end. The Vault closes that gap.
  • Where it runsHosted by Mnestiq in the EU, or on your own server, away from the agents. It stores the signed checkpoints, never the evidence itself. Its own key sits in a separate signing service, and the Workbench checks its log on every pass.
  • How it shows upA disagreement with the Vault raises a high-severity alert, in Sentinel too, and is stated in the claim summary.
For insurers and lawyers

A claim pack anyone can check.

  • The claim summaryOne printable page: what happened, when it started, when it was discovered and contained, what left and where, the agent's permissions and approvals at the time, how far the evidence can be relied on, and what would have prevented it.
  • Times set by peopleDiscovery and containment carry deadlines, so the analyst records them, with the time zone, in the chain of custody. Nothing is guessed.
  • The insurer packOne signed, redacted zip: the claim summary, the before-and-after page, the full claims report and the timeline. Personal details are masked throughout.
  • Checkable without usA VERIFY.txt lists every file's SHA-256 fingerprint, with the certutil and shasum commands to check them on any computer. Signatures can be checked with the Workbench.
Incident claim summary: what happened, when, what left, controls in place, evidence reliability and prevention
The claim summary from the incident above, as a claims handler receives it.
Chain of custody

Built to be cross-examined.

OriginalsEvery file is preserved byte for byte, named by its SHA-256, and rechecked on demand. The whole case can be rebuilt from the originals to show the analysis matches them.
Chain of custodyEvery ingest, decision, note, report and export is logged with who and when, each entry linked to the one before by its hash.
Signed outputReports, pages, timelines and bundles are signed with the case key, and each records the custody log's state when it was made.
RedactionCopies for outside readers remove prompt and document content and mask emails, URL paths, account numbers and local paths. Redacted copies are signed too.
Local by defaultThe web app listens on your own machine only and opens with a one-time key. Nothing is sent anywhere unless you configure a destination.
Hostile inputEvidence and logs are treated as untrusted: size limits, refusal of binary and oversized input, and evidence text is only ever displayed as text.
PackagedA Windows installer, a Debian or Ubuntu package, or a container image, compiled, with a self-check (mnestiq-wb doctor) that says what to fix.
Early access

Running agents with real permissions? Let's look at your first case together.

We're looking for a small number of security teams, MSSPs and insurers to shape the Workbench around real incidents. The recorder it builds on is open source.

alexander@mnestiq.com