The disclosure that reframed the argument
Anthropic reported in November 2025 that it had shut down a cyberespionage campaign in which its Claude Code tool performed 80% to 90% of the operational work. The actor, tracked as GTG-1002, kept people in the loop to approve decisions, but the model handled reconnaissance, exploitation attempts, and much of the documentation. That disclosure moved a long argument about AI and offense from speculation toward measurement.
The ransomware version of the same argument arrived in WIRED's reporting on what it describes as a new AI-fueled boom, in which incident responders say crews are folding models into lure writing, data triage, and tool development. The framing invites panic. The mechanics invite a narrower question: which parts of an intrusion are actually getting cheaper?
Where a model genuinely cuts cost
Fact: the labor in a ransomware operation is lopsided. Encrypting a network is a script that has existed for a decade. The expensive work sits at the edges: researching a target, writing a lure in the right language and the right register, sorting through terabytes of stolen files to find the contracts and the insurance policies, and drafting an extortion note that cites the specific regulator the victim fears.
Every one of those tasks is text-shaped, and text-shaped work is what large models do at near-zero marginal cost. A crew that once needed a fluent English speaker to hit a US manufacturer can localize a phishing thread into German, Portuguese, and Japanese overnight. The evidence on how far that goes is still thin: as I wrote when the published record on AI-designed malware was last assembled, the papers show assistance rather than invention. FunkSec, a group Check Point Research described in January 2025 as leaning heavily on AI-generated tooling, is the clearest public example of what a small, low-skill team can now assemble.
The part the models still fail at
Initial access remains stubbornly human. Verizon's 2025 Data Breach Investigations Report put the human element in about 60% of breaches and found ransomware present in 44%, and the entry points behind those numbers were mostly stolen credentials, commodity infostealers, and unpatched edge appliances. A chatbot cannot infer that a VPN concentrator three versions out of date is internet-facing. It can help write the scanner, but someone still has to run it and read the output.
Models also leave a smell. Agentic tooling generates API calls in patterns that detection vendors now fingerprint, and the same fluency that makes a lure convincing makes it uniform. The attacker's advantage in that exchange is throughput, and throughput is exactly what defenders can rate-limit.
And the affiliate economy itself has not dissolved. Chat rooms, access brokers, and negotiators are still people, which is why the hidden labor behind AI systems keeps reappearing inside crime as well as inside legitimate operations.
Attempts are up. Payments are not.
Fact: Chainalysis recorded roughly $814 million in ransomware payments in 2024, down from about $1.25 billion in 2023, a 35% decline. Interpretation: the two curves have separated. AI pushes the cost of trying toward zero, so attempts multiply and leak-site postings grow. Payment volume depends on something else, namely whether the victim has tested its restore process and whether insurers and counsel have learned to refuse.
The place this shows up first is the middle of the market: hospitals, school districts, logistics firms with 400 employees and one IT contractor. Those organizations have enough revenue to be worth extorting and not enough staff to rehearse. The operational fallout is the part the public sees, as it did when supermarket chains lost distribution systems and shelves went empty.
What to change, concretely
Retire the grammar test. The heuristic that a phishing email is clumsy and misspelled is dead, and any training program still teaching it is teaching the wrong lesson. Replace it with phishing-resistant multi-factor authentication on email, VPN, and identity providers: FIDO2 keys and passkeys, not push prompts that tired people approve at 11pm.
Measure recovery in hours, not in the mere existence of a backup. A restore that takes eleven days is a business-continuity event, and crews price their demands against the downtime they believe they can inflict. Patch the edge first, because the appliances terminating VPN and remote desktop traffic are where credential-free access actually happens.
None of that requires an AI strategy. It requires the work that security teams already know how to do and rarely get budget for: know your inventory, watch the logs, and keep the list of who holds admin rights short enough to read aloud.
What I expect next
Prediction: the next twelve months bring louder volume rather than new capability. Two parts are checkable by the end of 2026: leak-site postings and extortion-only cases rise, while publicly documented intrusions in which a model completed initial access with no human hands stay close to zero. Data theft, not encryption, drives the growth, because copying files needs neither domain admin nor a failed backup.
The regulatory clock is the underrated variable. US public companies have four business days to file an Item 1.05 8-K once they determine an incident is material, and CISA's CIRCIA reporting rule has slipped past its original deadlines. Cheaper attacks push more companies against that four-day wall. Prediction: by late 2026, incident descriptions in those filings will mention AI-assisted tooling often enough to be counted, and that count will carry more weight than any vendor survey about what attackers can do. If the filings show nothing of the kind, the AI-offense thesis is weaker than the reporting suggests.
Comments
No comments yet. Be the first to share your thoughts.
Leave a comment