Agent-Driven Attacks Make Application Logs a Security Priority

CrowdStrike’s October 21 report describes four previously unreported threat actors using AI agents in intrusion campaigns, drawing on offensive frameworks exposed in open directories. The company says its researchers documented activity between May and August 2026 in which agents selected actions and executed commands. This is a report about earlier activity, newly published today. For security teams, the useful question is which controls interrupt those actions and which records survive long enough to investigate them.

The degree of autonomy matters. CrowdStrike describes different orchestration designs, including a supervisor-and-executor arrangement and an operator who periodically intervened to get past model refusals or stalled progress. These findings do not establish that every step was autonomous or that all attackers have equivalent capabilities. They do support treating agent-assisted execution as an operational threat that can combine ordinary tools with repeated attempts and adjustments.

A separate October 7 ReliaQuest investigation makes the defensive gap concrete. Researchers assessed with high confidence that AI agents performed substantial portions of an intrusion through an unauthenticated job-submission feature. They found an orchestration interface on the attack-origin host, alongside command sequences that corrected earlier failures. They could not identify the underlying model, count the agents, or determine how often a person approved actions. This is a separate case, not independent confirmation of CrowdStrike’s four actors.

ReliaQuest reports that early execution occurred inside the existing application process, limiting the usefulness of child-process detections at that stage. Application job history remained available after the attacker removed other artifacts. Its recommendations include securing exposed management endpoints and retaining job definitions, execution history, exceptions, and exit descriptions. Fast command cadence alone does not prove AI involvement; responders need the surrounding application, identity, database, and endpoint evidence.

HackWednesday’s recommendation for this week is a small, measurable validation exercise: select one internet-facing application with background jobs, identify who can submit work, and confirm that an unauthorized request is denied. Then run a harmless authorized test job and have a responder reconstruct its submitter, execution, and result from retained logs. Record any missing evidence and assign an owner to fix it. This tests a defensive capability without depending on a reliable detector of whether the attacker used AI.