Adversaries may obtain access to generative AI tools, such as large language models, to help with various techniques during targeting. MITRE lists this under Resource Development with platform PRE. The current page was last modified in September 2026.
The technique is about access to the tool, not a particular attack. The tool is used to inform, bolster, and enable other malicious tasks.
How attackers use it
MITRE lists several uses. Adversaries may create scripts, assist social engineering, develop payloads, and query public AI services in support of reconnaissance.
By using a publicly available model, an adversary is outsourcing or automating tasks. MITRE gives examples: drafting content in many written languages for phishing and phishing for information, supporting vulnerability and other offensive research, and generating, refining, or obfuscating malicious scripts and payloads.
AI-generated text, images, audio, and video may also be used for fraud, impersonation, and other malicious activities.
None of this changes what the target sees at the end. A message still arrives, a script still runs, a call still asks for something. That is where defenders have visibility.
It is worth being exact about what MITRE claims. The technique is the obtaining of access to generative AI tools, and the examples describe tasks the tools can assist: scripts, social engineering, payloads, reconnaissance, and content for fraud. It does not claim that the tools create new attack classes.
The practical effect is on cost and quality. Writing a convincing message in a language the author does not speak, producing variations of a script, or generating audio or video for impersonation all become cheaper. The final behavior on your systems is still one you can describe, such as a message, a script, or a request.
MITRE also mentions querying public AI services in support of reconnaissance. A model can summarize public information about a target, which ties this technique to the reconnaissance techniques elsewhere in this series.
How to detect it
MITRE says much of this activity takes place outside the visibility of the target organization, so detecting the access to the AI tools is difficult.
Its guidance is to focus on behaviors related to the potential use of generative AI, naming phishing and phishing for information.
That suggests judging messages by what they ask for, not by how they read. Fluent, error free writing in any language is no longer a sign of legitimacy. Look at the request, the sender's domain and authentication results, links, and whether the ask bypasses a normal process.
For impersonation by audio or video, the signal is process: a request for money, access, or secrets that arrives through an unusual channel and creates urgency.
Here are three places to put detection effort. In email, look at sender authentication, new or lookalike domains, and requests that ask for money, access, or secrets. In endpoints, look at script execution and signs of obfuscation, since generated code still needs a runtime. In business processes, look at requests that skip normal approval.
Do not rely on trying to identify AI generated text. MITRE does not suggest it, and there is no dependable way to do it. A message is suspicious because of what it asks, who it claims to come from, and how it arrived.
Where voice or video is used, agree in advance on a challenge: a phrase, a callback to a known number, or a second approver for sensitive requests.
A last point on measurement. Track reports from staff about suspicious messages and calls, and note what the request was. Over a few months the list shows which pretexts reach your people, and that is more useful for training than any general description.
How to reduce the risk
- Be clear about the boundary. MITRE's own mitigation entry for this technique says it cannot be easily mitigated with preventive controls, because the behavior happens outside the scope of enterprise defenses and controls.
- Move verification out of band. For payment changes, credential resets, and access requests, confirm through a channel you already know, not the one the request came in on.
- Update awareness training. Remove advice that treats spelling mistakes as the main warning sign.
- Strengthen email authentication and filtering on your own domains and review reports of lookalike domains.
- Require phishing resistant authentication where you can, so a convincing lure alone does not give access.
- Keep the detection work on payload behavior, such as script execution and obfuscation, since generated code still has to run on your systems.
Treat this as a prompt to refresh basics. Update training examples, check email authentication on your domains, require strong MFA, and rehearse the out of band check.
If your own teams use generative AI tools, add a policy for what may be pasted into them. That is a separate risk from the one in this technique, but it belongs in the same conversation with staff.
Review the plan twice a year, because the tools change quickly. The controls that hold up are the ones that do not depend on how a message reads: verification, strong authentication, least privilege, and monitoring of what runs on your systems.
Check your own environment
- Does every high risk request (payments, resets, access) have a documented out of band check?
- Does your training still teach that bad grammar means phishing?
- Can staff report a suspicious message in one click, and does someone read the reports?
- Do you have a process for a request that arrives by voice or video from a senior person?
Quick check
What is the technique, in one sentence?
Obtaining access to generative AI tools to help with tasks such as scripts, social engineering, payload work, and reconnaissance.
Where does MITRE say detection should focus?
On behaviors related to the use of generative AI, such as phishing and phishing for information, because the tool access is invisible.
Sources
- MITRE ATT&CK, Artificial Intelligence (T1588.007): https://attack.mitre.org/techniques/T1588/007
