AI Agent Incident Readiness & Response

When an agent acts,
can you prove what happened?

An AI agent deletes data, pays the wrong vendor or sends the wrong message. The questions that follow come from the board, the insurer and the vendors. What did the agent see? What did it decide? What authorized it? We test whether your records can answer those questions before an incident. When one happens, we reconstruct it.

9
Times AI agents destroyed live
company systems, July 2025 – mid-2026 1
Jan 2026
ISO generative AI exclusions
available for general liability 2
$25M
Lloyd’s-backed AI liability limit
per organization, written against technical review 3
Why readiness, not just response

The evidence starts expiring before anyone asks for it.

An agent incident leaves records in several places: the agent’s harness, the model provider, each system the agent touched, and the configuration that told it what it could do. Each has its own owner and its own retention period, and some periods are short.

After the April 2026 PocketOS deletion, the most detailed public account of the agent’s reasoning was the agent’s own written reply. Two of the three companies involved gave competing accounts; the third said nothing. No neutral reconstruction was published, and attribution was never settled. Read the brief.

Carriers are drawing lines around this exposure. Some are excluding it; others write it affirmatively against evidence of controls. In both cases the question is the same: could the insured show what happened?

  1. StackGen, State of Reliability 2026: AI agents “have destroyed live company systems on their own at least nine times” since July 2025, each documented by an operator postmortem or incident database. StackGen is a vendor in this market. Release as reproduced by Continuity Insights, August 4, 2026.
  2. Verisk’s ISO introduced optional generative AI exclusion endorsements for commercial general liability (CG 40 47, CG 40 48) and products/completed operations (CG 35 08), edition date January 2026. Big “I”, October 21, 2025.
  3. Armilla Insurance Services, a Lloyd’s coverholder, raised its standalone AI liability limits to $25 million aggregate per organization in January 2026, pricing each system from a technical review. Armilla, January 2026.
Services

Three engagements. Each ends in a written report.

Could we reconstruct an incident tomorrow?

Readiness Review

We inventory your agents and the evidence each one leaves, run a 48-hour reconstruction drill on a real recent action, and score the result against the frameworks underwriters and auditors use (AIUC-1, NIST AI RMF, ISO/IEC 42001). Useful before a renewal or an AI liability application, and after major changes.

What you get
  • Agent inventory and evidence map
  • Drill results on one real action
  • Readiness score (0–100) mapped to AIUC-1, NIST AI RMF and ISO/IEC 42001
  • Gap report, ordered by cost to fix
  • Preservation playbook for your stack
How long
Typically two to three weeks from kickoff to report.
What we don’t do
Certify compliance with a standard, test model quality, or implement fixes in your systems.
Fee: $7,500–$12,000 fixed · details
An agent did something irreversible. What now?

Incident Response

We help you preserve the evidence in the first hours, then reconstruct the incident: what the agent saw, what it decided, what executed, and what authorized it. We attribute cause across deployer, harness vendor, model provider and tool vendors, as far as the record supports.

What you get
  • Preservation steps and vendor preservation requests
  • Hashed evidence set with custody log
  • Written findings report: timeline, causal chain, attribution, open questions
  • Remediation findings for the system
How long
Preservation starts on engagement. The reconstruction timeline is set at intake and depends on how quickly vendors produce records.
What we don’t do
Testify, act as an expert witness, provide litigation support, or work on contingency.
Fee: $3,500 preservation, then hourly · details
For carriers, coverholders and brokers

Claims Investigation

Panel investigation for losses involving AI agents, and pre-bind readiness assessments for applicants. We reconstruct the loss from records, identify the parties in the causal chain, and report what the evidence supports. The carrier uses the findings to make its own coverage and subrogation decisions.

What you get
  • Pre-bind: readiness score and gap summary for the applicant
  • Post-loss: reconstruction, causal chain and attribution findings
  • Records inventory showing what each party holds
How long
Pre-bind assessments typically two to three weeks. Post-loss timelines set with the carrier at assignment.
What we don’t do
Make coverage determinations, adjust claims, pursue recovery or subrogation, or provide expert testimony.
Fee: panel rates by agreement · details
Readiness Agent inventory · evidence map · score mapped to AIUC-1, NIST AI RMF and ISO/IEC 42001 · gap report · preservation playbook
Readiness SnapshotOne agent · questionnaire mapping and gap summary · no drill · about one week
$2,500–$3,500
Readiness ReviewUp to three agents · 48-hour reconstruction drill · two to three weeks
$7,500–$12,000
Each additional agent
$2,000
Annual re-reviewDrill re-run and re-score
50% of original fee
Incident Response Preservation, reconstruction and written findings
PreservationFirst 48 hours · retention freeze, exports, hashing, vendor preservation requests
$3,500
Reconstruction and findings reportBilled hourly against a retainer agreed at intake
$300–$350/hr
Claims Investigation
Pre-bind assessments and post-loss work for carriersHourly, under the carrier’s panel terms
By agreement
Partners
First referral from a new broker, carrier or implementer partnerApplies to the first Readiness Snapshot or Readiness Review each new partner refers
50% off
Brokers and underwriters who refer an applicant
No charge
Implementer white-label pricingImplementer resells at its own price
25–30% below list
How quotes work
Where a range is shown, the fee depends on the number of agents, the vendors in the chain and the systems they act on. We confirm scope and a fixed fee in writing after a short intake call, before work begins.
Who pays
The company whose agents are reviewed pays for readiness work. Referring brokers and underwriters pay nothing. Carriers pay for claims work they assign.
Our Agent Incident Process

The estate practice’s discipline, applied to a new kind of record.

The steps mirror our estate asset process: secure the record first, examine it without altering it, report only what it supports, and give it back.

  1. 01

    Intake

    A short call to understand the agents, the vendors in the chain, and the question to answer. If the matter is outside our scope, we say so. Scope, deliverables and fee are confirmed in an engagement letter.
  2. 02

    Preservation

    Retention is frozen on traces, audit logs and mailboxes. Records are exported and hashed at collection, with a custody log of who collected what, when and how. Preservation requests go to the model provider and each tool or infrastructure vendor. For readiness reviews, this step tests whether you could do the same.
  3. 03

    Reconstruction

    We establish what the agent saw, what it decided, what executed and what authorized it. We corroborate each step against records the agent could not write to, such as target-system audit logs and provider records. An agent’s own explanation is treated as a statement to test, not as evidence.
  4. 04

    Written Findings

    Every engagement closes with a written report. It sets out the methodology, the records examined and their hashes, the findings, the attribution the record supports, and the questions it cannot answer.
    • Readiness reviews: score, gap report and preservation playbook
    • Incidents: timeline, causal chain and attribution findings
    • Claims: findings the carrier can use to assess coverage and subrogation
  5. 05

    Return and Destruction

    Evidence sets and exports are returned to the client. Our working copies are securely destroyed at engagement close, consistent with NIST SP 800-88. The report and engagement records are retained as set out in our privacy policy.
Published analysis

Our method is public. Judge it before you call.

Incident briefs

Forensic reads of public agent incidents

What the record documents, what each party claims, what is missing, who holds the loss, and what a readiness posture would have changed.

Framework

Agent Incident Readiness Framework

Inventory template, evidence map, the 48-hour drill, gap fixes ordered by cost, a 0–100 score mapped to AIUC-1, NIST AI RMF and ISO/IEC 42001, and a first-hour preservation playbook.

About this practice

Who does the work.

The investigator, Matthew Snyder, has spent about 15 years in civil financial investigation and open-source research, including investigative work for a top-50 law firm. He also works hands-on with agent systems; he built PromptLock, a prompt-security and compliance product. About the firm.

We work for the company that deployed the agent, or for its broker, carrier or implementer, under a written engagement letter.

Contact

Book an intake call.

Tell us which agents you run, or what happened. The intake call is short. We will say whether a readiness review, incident response or claims engagement fits, and what it would cover.

If an incident is in progress: suspend the agent, do not delete it or its logs, and see the first-hour preservation playbook while you wait for our call.

Brokers, carriers and implementers: refer a client by email with their contact and a line on their agents. There is no cost to you.

We typically respond to email within one business day.

Email
Address
2345 Ashland Avenue, Suite 241
Cincinnati, Ohio 45206

This form does not send data to any server. It opens a drafted email in your own mail program, addressed to matthew@trustforensics.co. Do not include passwords, keys or credentials.