Skip to main content

DFIR for AI

AI Incident Response and Forensics

When an AI agent does something it shouldn't, you need to know what it did, what it was allowed to do, and why. We investigate incidents involving AI agent workflows — and help you make them investigable before something goes wrong.

Last reviewed: October 2026.

Prepare: Incident Readiness Review Urgent: contact us now

The short answer

AI Incident Response and Forensics is a scoped engagement to reconstruct an incident involving an AI agent: the inputs it received, the tools it called, the permissions it held, the decisions it made and the controls that did or did not stop it. It produces a written findings report with evidence, confidence levels and stated limits.

Two ways to engage

Planned work before an incident, urgent work during one. If you are not sure which you need, that itself is a readiness question.

Incident Readiness Review Urgent Triage and Forensic Review
When Before an incident During or after an incident
What you get An assessment of whether your workflow keeps the evidence an investigation needs (logs, model and prompt versions, tool calls, approvals), and a response plan First-response triage, evidence preservation guidance, and a forensic reconstruction with a findings report
Timing Scheduled, fixed fee Urgent engagement inside our coverage window; scoped and capped fee after triage

Fees are stated in the engagement letter for your market. Terms — including evidence handling, retention, deletion and any limitation of liability — are contractual, set out per engagement, and reviewed by counsel. The page deliberately does not summarise them.

Who it is for

  • Organisations running AI agents that read external input, call tools or touch customer, financial or operational data.
  • Teams that need to answer: what happened, how far did it reach, and could it happen again?
  • Risk, security, legal, compliance and operations leaders — the people who will have to explain the incident to someone else.

What we examine

Inputs

What the agent received, including retrieved documents and tool outputs, and whether any contained instructions.

Permissions

What the agent was authorised to do, and what it actually did (request versus grant).

Execution environment

What the sandbox or runtime permitted, and what the agent reached.

Decisions and approvals

Where a human approved — or should have.

Evidence quality

What logs exist, what is missing, and how confident each conclusion can be.

Control failures

Which controls held, which failed, and which were absent.

What you receive

  1. 1 Timeline of events with the evidence behind each entry.
  2. 2 Scope and impact assessment — what was reached, what was changed.
  3. 3 Control analysis: what held, what failed.
  4. 4 Confidence statement per finding, with gaps stated.
  5. 5 Recommendations for containment, recovery and control changes.
  6. 6 A handover pack your counsel, insurer or regulator-facing team can use.

Findings can be relied on by third parties — insurers, regulators, litigants, law enforcement — which is exactly why each one carries its confidence level and its limits, and why the report is written to be handed over, not just presented.

What we do not do

We do not provide 24/7 or on-call cover outside our stated hours. We do not guarantee a root cause or attribution: AI systems are not deterministic and evidence may be incomplete. We do not provide legal opinions, regulatory filings or expert-witness testimony unless separately agreed. We do not take over containment or remediation unless agreed in writing — your own team or an incident response firm remains responsible for containment unless we have scoped that work. If your situation is an active, large-scale breach beyond our capacity, we will say so and refer you to [VERIFY: the referral partner for cases beyond our capacity] .

Urgent contact

Live incident? Contact [VERIFY: the monitored urgent channel — an address or number that alerts the principal] . Coverage: [VERIFY: hours and time zones actually covered — state them exactly; do not imply round-the-clock service] . First response: [VERIFY: a tested first-response time inside the coverage window] . Outside these hours, contact your incident response provider; we will respond when the window opens. This route is monitored by the principal. The general contact form is not an emergency line.

We publish a coverage window rather than a 24/7 promise because support needs vary widely depending on the incident.

How an engagement runs

  1. 1

    Engagement letter

    Before evidence is accessed: scope, confidentiality, evidence handling, terms. A short engagement letter precedes any work — it protects both sides, and we say so plainly.

  2. 2

    Triage

    Establish the situation, preserve evidence, and decide whether the work fits — including telling you when it does not.

  3. 3

    Reconstruction

    Analyse logs, prompts, tool calls, approvals and the execution environment against the timeline.

  4. 4

    Findings

    A written report with confidence levels and stated limits per finding.

  5. 5

    Debrief

    With your team, counsel and, where relevant, your insurer.

Evidence stays in an access-controlled workspace with a chain-of-custody log; retention and deletion are agreed in writing. Handling of personal data inside evidence needs a lawful basis and a data-processing agreement in each market — we raise this at the engagement-letter stage, not after.

Why Banbreach

Banbreach's principal has worked on data breach, ransomware and digital fraud cases : [VERIFY: scope of prior breach/ransomware/fraud casework, with confidentiality cleared] . [VERIFY: principal name] leads incident work personally. [VERIFY: N years of experience including sandbox-evasion analysis — scope and period stated] . Our product background means we also ask the question responders skip: how should the system have been built so this could be detected and stopped?

Meet the principal

Where this fits

If you have not had an incident, start with the AI Workflow Assessment, which includes the threat model this work depends on. If you want leadership alignment first, the Executive Workshop covers authority and escalation. The methods behind this work — including what forensic analysis can and cannot establish — are published on Methods & Demonstrations.

Enquiring about this work

Planned enquiry — readiness review or a question about method: use the general form and pick the incident category. Live incident: use the urgent channel. The form is not an emergency line.

Go to the enquiry form