How are people who misuse an AI model for cyberattacks actually using it, and do the measures that threat analysts use to rank attackers still work when a model does part of the work? This report from Anthropic's Frontier Red Team answers from Anthropic's own enforcement records. It takes 832 accounts banned for malicious cyber activity over one year and maps what they did onto MITRE ATT&CK, the security field's public catalog of attacker tactics and techniques. It finds that most actors used Claude to prepare attacks rather than to carry them out. It also finds that the actors scored as most dangerous are distinguished less by skill or breadth than by using the model inside compromised networks and by building agentic scaffolding around it.
The report was published on anthropic.com on 3 June 2026 under the Frontier Red Team label, by Kyla Guru, Alex Moix and Jacob Klein (byline), with the title "Mapping AI-enabled cyber threats: Insights from the LLM ATT&CK Navigator." It first appeared on red.anthropic.com (the companion post links it there), and the Wayback Machine's index lists a first capture of that address on 3 June 2026; the captures themselves could not be retrieved for this note, so the text of the early red.anthropic.com version was not compared with the current page. That address now redirects to anthropic.com. A shorter companion post on the same day, "What we learned mapping a year's worth of AI-enabled cyber threats," summarizes it and adds a few figures. Some results also appear in Verizon's 2026 Data Breach Investigations Report (DBIR), which used 11 months of the data (footnote 1); the DBIR was not read for this note. The report links an interactive Navigator page; its data are loaded by script and were not read here. The report's five charts print their values, and numbers below come from them where the text gives none.
Terms used in this note
- MITRE ATT&CK: a public framework that names attacker goals as tactics (14 in the enterprise matrix, such as reconnaissance, defense evasion, lateral movement and exfiltration) and the means to them as techniques and sub-techniques, each with an ID such as T1587.001 (Develop Capabilities: Malware). The report uses version 18.
- TTPs: tactics, techniques and procedures.
- Kill chain: the sequence of stages in an attack, from reconnaissance to impact.
- Lateral movement: moving from one compromised machine to others inside a network.
- ARiES (AI Risk Enablement Score): the report's new 0 to 100 score for how concerning an AI-involved misuse case is (defined below).
- Scaffolding: in the report's words, "the surrounding code, architecture, and tooling" built around a model so that it can chain attack stages with little human input.
The data and how they were coded
The sample is 832 accounts banned between March 2025 and March 2026 for violating the cyber provisions of Anthropic's Usage Policy (section "About the dataset"). Anthropic found them through automated safeguards and investigations by its Threat Intelligence team. The report says they are "just a subset" of the accounts banned in that period, chosen because there was enough detail to map their techniques (section "Key findings"); it does not give the total banned.
For each account, a summary of observed activity was produced, and TTPs were extracted from the summaries and mapped to ATT&CK v18. The report does not say who or what wrote the summaries or did the extraction, or whether any of it was checked by a second coder. In all it records 13,873 observations across 482 unique techniques (the report says "sub-techniques" in one place and "techniques" in another) and all 14 tactics. Enterprise techniques make up 99% of observations, the rest from ATT&CK's mobile matrix (footnote 2). The data were anonymized.
The ARiES risk score
ARiES adds three components (section "The LLM ATT&CK Navigator and ARiES risk score"):
- Threat (0 to 35): clarity of intent, technical sophistication, threat-intelligence signals and evasion. Technical sophistication "is graded by Claude on the basis of the actor's prompts and tool usage."
- Vulnerability (0 to 35): the model's capacity to enable the requested harm and the risk of the interface. The API and agentic coding tools such as Claude Code "score highest."
- Impact (0 to 30): scores from Anthropic's safety classifiers and investigators' assessment of actual or potential consequences attributable to the AI's involvement.
Totals are binned into low, medium, high and critical tiers; the cut-offs are not given. The report explains why it adds rather than multiplies (section "A note on the scoring formula"). A product of threat, vulnerability and impact falls to zero when any one is absent. The authors want a score that still flags, for example, an inexperienced user who produces a working offensive tool without clear intent, or a malicious actor whose malware has no known victim yet. They state the cost: the scores "are not predictions of whether an attack will be successful," but measures of how concerning a case is. No validation of the score against outcomes or against independent analysts' rankings is reported.
What it finds
Most use is preparatory
- Capability development (T1587) was the most common technique family, used by 574 of 832 actors (69%), mostly as malware development (T1587.001, 560 actors; 67.3% in the companion post). The next most prevalent were obfuscation (T1027, 64.7% of actors), data from the local system (T1005, 55.9%) and impairing defenses (T1562, 54.9% in one place and 54.8% in another) (section "AI-assisted capability development").
- Defense evasion was the largest tactic: present for 84.4% of actors, and 2,243 of 13,873 observations (16.17%) (section "AI-assisted evasion tactics"; tactic chart). The report observes 32 of the 64 defense-evasion techniques in ATT&CK's enterprise and mobile matrices.
- Later stages were rare. Impact and exfiltration each made up 2.78% of observations, privilege escalation 2.36% and lateral movement 0.7% (97 observations) (tactic chart); the text gives their total as 8.7%, the sum of the rounded shares (the chart's counts give 1,196 of 13,873, or 8.6%, our arithmetic). Only 54 actors (6.5%) used the model for lateral movement, and fewer than 12 for remote services such as SSH and SMB.
- Breadth per actor. The median actor used 16 techniques (mean 16.26) and 7 tactics (mean 7.08) (coverage chart). The report notes that this breadth "five years ago, may have signaled a well-resourced, technically mature operation."
The top techniques changed little over the year, and the median stayed at 16 in both halves. In the second half the report sees a "subtle directional shift" toward use inside networks: account discovery (T1087) rose 8.9% and automated exfiltration (T1020) 6.2%, while capability development fell 12% and phishing 8.6%. The report does not say whether these are changes in observation counts or in shares of actors, or whether they are relative or in percentage points.
What marks the highest-scoring actors
- Lateral movement. The 54 actors who used the model for lateral movement averaged 56.44, against an overall mean of 46.76 (chart "Tactics by mean actor risk score"). The text says they score "on average, 10.5 points higher" than actors who did not, and that "No other technique came close to having such predictive power." From the chart's means and counts, the other 778 actors average about 46.1, a difference of about 10.4 points (our arithmetic). Every other tactic's mean in the chart lies between 44.51 (privilege escalation) and 49.24 (exfiltration). The companion post lists privilege escalation, with account discovery and lateral movement, among the techniques on which higher-risk actors "concentrate their use of AI"; in the report's chart privilege escalation has the lowest mean of any tactic (this note's comparison).
- Post-compromise techniques. Remote services, valid accounts, credential dumping, archiving collected data and web shells were each "three to five times more common" among the highest-risk actors than in the whole population (section "High-risk actors and their tactics").
- Weak predictors. Technical sophistication, once removed from the composite, correlated with the remaining components at r = 0.28. Removing it left the top six actors in the same order (Spearman ρ = 0.96 across all 832). Breadth of techniques correlated with risk at r = 0.27. 80% of actors misused Claude Code, and the report says actors restricted to the chat interface, the API or agentic tools had "statistically indistinguishable risk profiles," without reporting the test. The companion post adds that the least-skilled actors used about 16 techniques on average and the most skilled about 20.
The report reads these results as saying that the actors who get the most from AI "simply used Claude for more hands-on techniques."
The rise in risk scores
The share of actors at medium risk or higher rose from 33% (33.5% in a later section) in the first half to 56% (56.1%), which the report calls "a 1.7x increase in under a year" (sections "Key findings" and "Live exploitation actors on the rise"). The risk-distribution chart gives the counts behind this:
| Period (chart labels) | Accounts | Low | Medium | High | Critical |
|---|---|---|---|---|---|
| Mar to Sep 2025 | 529 | 352 | 173 | 2 | 2 |
| Oct 2025 to Mar 2026 | 303 | 133 | 167 | 2 | 1 |
Data from the report's chart "Risk category distribution." The chart's period labels span seven and six calendar months; the text calls both "six-month" periods. From these counts, 177 accounts were at medium or higher in the first half and 170 in the second (our arithmetic). The rise in share therefore comes from fewer low-risk accounts in the second-half sample, with the number at medium or higher about the same. High and critical together were 4 and 3 accounts. The report's first finding is headed "The number of actors using AI for cyber operations is growing," but the analysis covers a selected subset, and the counts it shows fell between the halves. The report allows that "improved threat detection techniques may have contributed to this increase," and also points to more actors asking for in-network work.
One actor: GTG-1002
The report scores the actor behind the cyber-espionage campaign Anthropic disclosed in November 2025, labeled GTG-1002, at the maximum of 100 (section "Novelty and sophistication in the age of AI agents"). Its ATT&CK profile, 30 techniques across 13 tactics, was comparable to "dozens of medium-risk actors"; several low-risk actors also exceeded 30 techniques. The report attributes the score to orchestration. The actor ran Claude Code with open-source penetration-testing tools exposed as MCP (Model Context Protocol) servers, so that the model executed reconnaissance, exploitation, credential harvesting and lateral movement itself. The human operator gave strategic direction and kept control of the final data extraction. The report's conclusion is that "a technique-frequency table cannot capture" this. The case is a single actor, scored with the same method as the rest.
What the report proposes
- Detection. Anthropic says it has updated classifiers and expanded its probe detections to catch techniques that correlate with high ARiES scores, and is building signals for agentic patterns that ATT&CK does not name (section "How we are using the Navigator to inform our safeguards"). It reports no measurements of these changes.
- Access controls. Real-time cyber safeguards on its most capable models block prohibited activity at the request level, and higher-risk dual-use work is routed through a Cyber Verification Program for vetted defenders.
- The taxonomy. Autonomous kill-chain orchestration, real-time pivot decisions and AI-directed execution "don't yet have ID numbers in the ATT&CK framework." The report proposes cross-cutting ATT&CK categories for such behavior and says Anthropic is in "active conversations with MITRE" (section "A new era for MITRE ATT&CK").
The written framework in this report
The report contains no principles or prompts given to a model. The written texts that shape its results are classification schemes: ATT&CK v18, whose categories every observation was mapped to, and the ARiES rubric with its point ranges. One component of the rubric, technical sophistication, is applied by Claude. The report's argument that ATT&CK lacks categories for agentic behavior is an argument about a written taxonomy's coverage. The report says all 13,873 observations "mapped to categories laid out in the framework" (section "Key findings"). On this note's reading, the agentic behaviors the report says are missing therefore do not appear in its counts, and the evidence for them is its account of cases such as GTG-1002.
Limitations and open questions
- Selection. The sample is banned accounts with enough detail to map, a subset of an unstated total. It reflects what Anthropic detected and documented, so changes over time can reflect detection as well as behavior, as the report itself allows.
- Unvalidated coding. Who wrote the activity summaries and extracted the TTPs is not stated, and no agreement between coders is reported. Technical sophistication is graded by Claude without reported validation.
- The score's construction. The report's first finding says the techniques that grew, such as lateral movement and credential dumping, "carry the highest per-actor risk weight in our scoring." On this note's reading, their association with high scores is therefore partly built into ARiES. The report removes technical sophistication to check for circularity but reports no similar check for technique weights. The interface also enters the Vulnerability component, yet the report finds risk profiles indistinguishable across interfaces; with 80% of actors on Claude Code, the comparison groups are unequal in size.
- Reporting gaps. The tier cut-offs, the test behind "statistically indistinguishable," and the units of the technique changes (8.9%, 6.2%, 12%, 8.6%) are not given. Small inconsistencies in the text (33% and 33.5%, 54.9% and 54.8%) do not change any conclusion.
- One provider. The data cover only Claude's products. Actors who were not banned, or who used other models, are outside the sample.
- The publisher's stake. The report is a developer's account of misuse of its own products and of its responses to it. Its proposals (new ATT&CK categories, routing dual-use work through Anthropic's verification program) are positions, not results.
Research questions
- Does ARiES rank actors as independent analysts would? Having analysts outside the team rank a sample of accounts blind to the score, and measuring agreement, would test the score's validity.
- How much of the lateral-movement signal survives reweighting? Recomputing ARiES with equal technique weights, or without the techniques of interest, would show whether their association with high scores is more than the weighting.
- What does the full population of banned accounts look like? Reporting the total banned per half and the share that could be mapped would show whether the fall from 529 to 303 accounts reflects fewer actors, fewer detailed records or different detection.
- How reliable is the TTP extraction? Double-coding a sample of activity summaries against ATT&CK, with an agreement statistic, would show how much noise the counts carry.
- Can orchestration be measured? Defining observable indicators of agentic scaffolding (tool-call chains, MCP servers, turns without human input) and counting them across the 832 accounts would test the report's claim that scaffolding is the differentiator, beyond the GTG-1002 case.
- Do the classifier and probe updates catch high-ARiES actors earlier? Comparing time to detection before and after the updates would measure the effect the report describes.
How it connects
- Claude's safeguards. The report links its probe detections to Constitutional Classifiers++; the classifier approach began with Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming.
- Capability, not misuse. The report sets its findings beside Claude Mythos Preview, which it describes as showing "where frontier AI cyber capabilities are heading." See Assessing Claude Mythos Preview's cybersecurity capabilities and the Claude Mythos Preview system card. The Risk Report: August 2026 is Anthropic's later company-wide assessment of cyber and other risks (this note's connection).
- Open-weight models. GLM-5.3 and the spread of advanced cyber capabilities (September 2026) is the Frontier Red Team's later analysis of a model without such safeguards (this note's connection).
- Cyber defense. Building AI for cyber defenders (September 2025) cites earlier misuse cases from within this report's window, and Experimenting with AI to defend critical infrastructure (January 2026) describes Claude used for defensive adversary emulation (this note's connection for both).
References
- Guru, K., Moix, A., and Klein, J. (2026). Mapping AI-enabled cyber threats: Insights from the LLM ATT&CK Navigator. Anthropic, 3 June 2026. https://www.anthropic.com/research/attack-navigator record
- Anthropic. (2026, June 3). What we learned mapping a year's worth of AI-enabled cyber threats. https://www.anthropic.com/news/AI-enabled-cyber-threats-mitre-attack record
- MITRE. ATT&CK, version 18. https://attack.mitre.org/versions/v18/. Cited in the report; not read for this note.
- Verizon. (2026). 2026 Data Breach Investigations Report. https://www.verizon.com/business/resources/reports/dbir/. Cited in the report; not read for this note. record