Query Public AI Services (T1682) is a Reconnaissance technique. Adversaries ask publicly accessible AI services, such as large language models (LLMs), to help with targeting and operations. MITRE lists PRE as the platform, since it happens before any attack on your systems.
How attackers use it
Adversaries already search websites and databases directly, which MITRE covers under Search Open Websites/Domains (T1593). AI services add the ability to synthesize, aggregate and analyze public information at scale.
MITRE lists the uses:
- Identifying people or organizations to target.
- Researching organizational structure and personnel.
- Identifying technologies a target uses.
- Researching business relationships to build believable pretexts for Social Engineering (T1684).
- Finding contact details for Phishing (T1566) or Phishing for Information (T1598).
- Gathering sensitive or derogatory information about individuals for extortion or coercion.
What is learned this way can feed other behavior, such as Generate Content (T1683) and Establish Accounts (T1585). Documented examples include groups that used LLMs to search for official email addresses, build target lists, identify think tanks, government bodies and experts, and research possible business partners.
How to detect it
MITRE is direct about this. Much of the activity takes place outside the target's visibility, so detection is difficult. The suggestion is to focus on later stages of the adversary lifecycle, such as Initial Access.
How to reduce the risk
The Pre-compromise mitigation applies: this behavior cannot be easily mitigated with preventive controls because it happens outside enterprise defenses.
What you can control is how much useful material is out there and how well your people and processes hold up when someone uses it:
- Review what your organization publishes about staff, structure and technology, and trim what does not need to be public.
- Train staff to expect messages that are accurate and well researched. Accuracy is no longer proof of legitimacy.
- Use verification steps for payments, password resets and access changes, so a convincing pretext is not enough.
- Strengthen the detections that sit at Initial Access, which is where MITRE points.
Quick check
Why is this technique hard to detect?
The adversary queries a public AI service, outside your network and logs. You see nothing until a later stage, such as a phishing attempt.
Sources
- MITRE ATT&CK, Query Public AI Services (T1682): https://attack.mitre.org/techniques/T1682
