Your company does not need AI software to suffer an AI-assisted attack
The affected system may be an ordinary Microsoft 365 mailbox, a Windows laptop, a VPN account or a finance process. The attacker may have used AI to write a convincing message, imitate a voice or accelerate preparation. Your business can therefore need AI incident response without running a chatbot, model or autonomous agent of its own.
The urgent question is what the attacker accessed or changed. A polished phishing message is a clue to investigate, not proof of AI use. Do not delay containment while debating how the lure was written. Record the observation and respond to the compromised identity, device, application or transaction. Attribution can remain uncertain while practical response continues.
Separate three different meanings of AI incident response
| Meaning | Example | Response priority |
|---|---|---|
| An attacker uses AI | A convincing impersonation leads to a stolen account | Investigate the account and business impact |
| Defenders use AI assistance | An analyst uses approved tooling to organise evidence | Validate outputs against original records |
| An AI system is compromised | An agent takes an unauthorised tool action | Investigate the AI workflow and connected systems |
These categories can overlap. The important change in scope is to include the systems your company actually depends on. A conventional mailbox investigation should not be replaced by a prompt-injection checklist simply because the attacker may have used an LLM.
Build an initial incident picture from observations
Ask the reporting person what happened, when, on which device or account and what they did next. Did they enter credentials, approve a sign-in, install software, send a file or authorise a payment? Keep their original message and timestamps. An employee should not be asked to reopen a suspicious attachment to demonstrate the problem.
Name the incident owner and identify a communication route the attacker is unlikely to see. Separate ongoing harm from historical exposure. Unauthorised payments, active encryption and suspicious mailbox forwarding create different immediate decisions. For each decision, record who can approve the change and which service could be disrupted.
Use AI to assist analysis without handing it operational authority
Approved AI tooling can help organise a large set of observations, draft a timeline or suggest questions for an analyst to test. That is assistance, not a confirmed finding. Each important assertion should point back to a source record. A fluent summary must not become the only surviving version of the evidence.
Agree whether evidence may be processed by an AI service, what data is sent and which retention and access controls apply. Keep credentials and unnecessary personal data out of prompts. Treat attacker-authored emails and documents as untrusted input, including when an assistant analyses them. Containment and recovery actions remain subject to human approval.
Ask for a response plan tied to your environment
A useful scope names the tenant, endpoints, servers, identities and business processes involved. Ask what records are available, what access the team needs and what happens if logs have expired. The proposal should distinguish initial triage, forensic investigation, remediation and recovery validation, rather than sell an undefined promise that AI will solve the incident.
Microsoft has published an investigation of a phishing campaign it assessed as likely using AI-generated obfuscation. That demonstrates a real use case, but it does not justify labelling every suspicious message an AI attack. The appropriate response remains evidence-led: identify access, contain ongoing harm and verify the repaired boundary.
Get help with the affected business systems
Explore this response workstream or our AI incident response services. You do not need to operate an AI application or prove the attacker used AI to request help. Describe the affected systems, observed impact and urgency; we confirm availability and agree an authorised scope.
Send only non-sensitive context through the website. If company email may be compromised, use a trusted communication channel and avoid sharing the investigation through the suspected account.
Sources and review
Reviewed 24 October 2026. The decision tables and planning examples are Atlant Security guidance. A suspected use of AI is not established attribution. Scope, availability and evidence handling are agreed before an engagement.
