Trends drawn from every verified case file on this site, recomputed each time the registry updates. The charts measure what has been publicly reported and met our inclusion rule, not the total amount of AI-enabled crime.
318Verified case files
31This quarter · 2026-Q4so far · previous quarter: 41
42%Target the AI stack itself135 case files
$225MDocumented losses31 files disclose a loss
54%Still ongoing or under investigation
9Attacks by autonomous AI agents
Attacks on AI are catching up with attacks with AI
Attacks on AIAttacks with AI
Verified case files per quarter, split by direction. Attacks on AI systems (models, agents, AI toolchains) were 35% of files in the first half of 2025 and 52% in the latest half-year. The AI stack became a target class in about a year.
Excluding researcher demonstrations: 270.
Which attack types are rising
Each cell counts the verified case files of one type in one quarter. AI supply chain has grown most over the last four quarters; Deepfake fraud remains the largest category overall with 82 files.
The rise of the autonomous attacker
Case files in which the attacker was an AI agent acting without a human instruction to attack: 2 in 2025, 7 in 2026. Most began inside AI labs' own evaluations and broke out of them. These are the files:
CriminalState-sponsoredSecurity researchersUnknownAutonomous AI agentAccidental / misconfigurationHacktivistInsider
Criminals account for 48% of case files. Researcher demonstrations (48 files) are counted and labelled rather than mixed in; state actors (38) are the third group; the attacker is unknown in 55 files, a share that is growing as reporting outpaces investigation.
Are incidents getting worse?
CriticalHighMediumLow
Severity mix per quarter. 57% of all case files are rated critical or high: confirmed intrusions, material fraud losses, active exploitation or state operations.
Where the money goes
Deepfake fraud (29)$224M
AI supply chain (1)$500K
LLMjacking (1)$600K
Documented losses total $225M, but only 31 of 318 case files disclose a figure, and 100% of the total is deepfake fraud. Technical breaches of AI systems almost never come with a published cost, so this chart understates them.
Who gets hit, by sector
Technology & AI vendors130
Individuals & consumers84
Multiple / cross-sector40
Unknown18
Government & public sector12
Financial services11
Defence & security9
Politics & elections4
Education & research3
Media & entertainment3
Energy, industrial & infrastructure2
Retail & consumer brands2
Sector of the victim, tagged by the editorial team for every case file. Technology and AI vendors absorb the attacks on AI systems; individuals and financial services absorb most of the fraud carried out with AI.
44% of case files are cross-border campaigns with no single country. Of the rest, United States leads with 67 files. Country counts reflect where incidents are reported, so English-language reporting is over-represented.
How it ends, and how fast we file it
Under investigation81
Ongoing74
Contained62
Patched54
Resolved30
Unresolved17
54% of case files are still ongoing or under investigation at the time of their last update. Median time from a story's first appearance on the wire to a verified case file, for files that started as robot drafts: 20 hours.
Stories per day on the wire, from more than fifty sources, by kind as sorted by the robot. Research and vulnerability stories dominate the conversation; confirmed incidents are a small share of what gets published.
Based on AI Threat Watch's internal analytics of 318 verified case files as of Oct 15, 2026. Trends start in 2025, when systematic coverage began; three earlier cases are counted in totals only. Researcher demonstrations are included and labelled. Severity and category follow the method described on the About page. Numbers: .Numbers as JSON. Charts are inline SVG: copy them into an article with a credit to AI Threat Watch (CC BY 4.0).