Verizon's 2026 breach report puts the median AI-assisted actor at about 15 known ATT&CK techniques

The two figures together describe a narrow population. A median actor sought AI help with roughly 15 distinct MITRE ATT&CK techniques, and fewer than 2.5% of the techniques sought were rare ^[4][5]. At that rate a rare one turns up about once every 40 requests: 1 divided by 0.025 is 40, which is about 2.7 times the median list of 15 ^[14]. The expected number of rare techniques on a median list is therefore under one, at 0.375 ^[15]. The 2.5% is a share measured across the population, so the per-actor figure is an approximation.

Actor profiling pays for itself when a group's name predicts which slice of the catalog you will see. Baker's argument is that AI severs the prediction and leaves the slice in place. "The predictability we lost was in the human," he wrote ^[8]. His supporting evidence is Anthropic's published mapping of a year of AI-enabled attacks, in which every observed behavior fit an existing entry in the framework ^[6].

Three caveats sit on this. The DBIR is subscription-gated ^[3], the summary comes from a vendor executive ^[1], and Anthropic supplied data for the analysis ^[6]. They weigh on the recommendation, and less on the number. The claim that AI produced no new techniques is falsifiable inside a defender's own telemetry, and a single observed behavior that does not map to ATT&CK breaks it.

The defensive case then rests on where paths converge. Baker points to chokepoints, using process injection as the example: many techniques lead into it and many follow from it, so defending that one technique disrupts many attack paths at once ^[9]. "AI does not move the chokepoints," he wrote ^[10].

The second half of his argument is about what a detection is anchored to. MITRE's Summiting the Pyramid methodology scores a detection on robustness against evasion, so a detection tied to a value the adversary changes at will, such as a file hash or a command string, scores low, and one tied to the behavior at the core of the technique scores high ^[11]. "A detection that survives new procedures is worth more than 10 that a single new one evades," Baker wrote ^[12].

The trade-off is between two kinds of spending. Validation against a fixed technique list is a program a security team can audit internally; group-specific playbooks are intelligence someone else maintains and sells on renewal. Baker's recommended moves stay inside detection engineering: assess defenses against the chokepoints and invariants, and prioritize coverage by terrain, business context and chokepoints ^[13]. The post confines itself to that question, and leaves aside whatever a group's name still buys elsewhere in a security program. For a team budgeting this quarter, the comparison is between maintaining actor-specific playbooks and validating coverage against the roughly 15 techniques a median actor reaches for ^[4].