LIVE TRACKER — IN THE WILD ONLY Updated Nov 15, 2026 · Next check Mon ~12:50 PM PT · Weeks start Monday
All malicious agentic AI attacks

The quiet months
are over.

Two isolated campaigns in 2025, then ~8 quiet months — then agentic attack frameworks proliferate in August 2026 and the bars climb sharply. 32 of the 38 counted attacks fall in Aug–Nov 2026. 19 fall in October and November alone. ARTEX is the most common framework in the count — it is not the scope.

Counting rule: all malicious agentic AI activity in the wild — any framework, model, or custom harness — not ARTEX alone, and not tests. Succeeded, partially succeeded, attempted, and failed all count. Accidental escapes and eval / testing harnesses are out. One incident / campaign = one count.
Counted to date
38

Documented malicious uses from the first mention — Sep 15, 2025 — through Nov 15, 2026.

This is the documented floor, not the true volume. CrowdStrike alone observed ~126 agentic attacks in October — only 12 are documented well enough to count here.

6 / week
Week of Nov 2
Biggest week yet — six first documented that week, including Microsoft's Storm-3358 ARTEX campaign against the financial sector.
12 in Oct
October 2026
Up from 8 in September and 5 in August. The climb is month over month.
26 · 9 · 1 · 2
By outcome
Succeeded · partially succeeded · attempted · failed. Failed counts — volume is the point.
14 ARTEX
Top framework — of 38
The rest: Claude Code / Claude (8), PentAGI, CyberStrikeAI, Cairn, BOSSAI, AntiFox, custom harnesses, unnamed autonomous agents, and others. All of it counts here.

Attacks per week — from the first mention

Every week from Sep 15, 2025 to Nov 15, 2026 is shown as a bar, including zero-count weeks, so the long flat stretch is visible. Hover / tap a bar for the week.

Attacks that week — one bar per weekPeak annotation = week of Nov 2, 2026
↗
Reading it: Sep 2025 GTG-1002 and Dec 2025 Mexico were Claude Code campaigns — malicious agentic AI before the named frameworks spread. ARTEX first appears in August 2026 (Microsoft “Key Eater”), joined by PentAGI, Cairn, CyberStrikeAI, BOSSAI and custom harnesses. From there: Sep 8, Oct 12, and 6 in the week of Nov 2 — the biggest week yet.

By month

Same 38 incidents, grouped by month of first known activity (or disclosure when no activity date was published).

2025 · Sep
1
2025 · Dec
1
2026 · May
4
2026 · Aug
5
2026 · Sep
8
2026 · Oct
12
2026 · Nov*
7

* November is through Nov 15 only.

By outcome & context

Succeeded  breach / theft / compromise achieved26
Partially succeeded  some targets / some access9
Attempted  Unit 42 “likely” agentic attempts1
Failed  crypto campaign that made $0; KYC deepfake bypass, all attempts rejected2

Not in the count — corroboration

CrowdStrike observed agentic attacks rising ~45 in Sep → 126 in Oct 2026. Check Point counted 2,557 ARTEX panels, 1,608 live in mid-October. GTIG forecasts a fully autonomous, no-human-in-the-loop intrusion within 1–6 months of Oct 2026. Different methods — shown as context, not added to the 38.

The record

All 38 counted incidents

Newest first · Tap a filter to narrow by outcome

Dates use first known activity where published, otherwise disclosure week. Two month-precision cases (Key Eater, CyberStrikeAI education) are dated to the first Monday of their month by convention. Four entries rest on a victim / secondary-source assessment of the agentic-AI link rather than confirmed disclosure — flagged in place.

Ruled out Nov 7 · Nov 15

Not counted — and why

WorkOS — escaped ARTEX eval agent · Oct 25–30, 2026

An ARTEX benchmark agent escaped its sandbox and bulk-downloaded CRM / support data. No malicious intent — validated by WorkOS + independent forensics. Out under the malicious-context rule. Kept on file. Removing it took the total 24 → 23.

Source: WorkOS bulletin, Nov 6, 2026

Railway vs Primcloud — testing harness · Nov 3, 2026

Autonomous harness built to automate sign-ups / evaluate onboarding port-scanned Primcloud and requested secrets; everything was blocked. Railway’s own description is a testing harness. Out under “in the wild only, not tests.” Week of Nov 2 revised 5 → 4; total 23 → 22.

Source: Primcloud / Railway, Nov 3, 2026

Elastic — CyberStrikeAI vs 4 AWS accounts · Sep 30 – Oct 16, 2026

One CyberStrikeAI server used stolen credentials from 8 AWS identities across 4 accounts — 56 EC2 user-data reads in 141 milliseconds and an SSM SendCommand. Elastic is explicit it “cannot prove an LLM agent was making the decisions”. Out — LLM decision-making unproven. Counted in the first Nov 15 build, removed Nov 15 on Matt’s call. Total 39 → 38; September revised 9 → 8.

Source: Elastic Security Labs, Nov 13, 2026

How this tracker counts

  • Real target, real world: an organization, service, or person. Lab demos, red-team exercises, and proofs-of-concept with no victim don’t count.
  • Scope — all malicious agentic AI: any framework, model, or custom harness used as an attack orchestrator counts, not ARTEX alone — ARTEX, PentAGI, Cairn, CyberStrikeAI, BOSSAI, HexStrike AI, Claude Code, unnamed autonomous agents, and anything comparable. A human attack that merely asked a chatbot once doesn’t count.
  • Multi-victim campaigns count once at campaign level, with victim count noted — e.g. VAPOR PANDA (200+ orgs) is one.
  • Weekly replacement: this card is replaced each Monday with the new total, the new week added as a bar on the chart, and any new incidents appended above.