Artificial Intelligence Policy Tracker

AI risk

AI attacks and incidents: the record in charts

What has gone wrong with AI, who caused it, whether it was meant, and how that compares with the risks that frameworks describe. Every figure is a count of records in the AI Incident Database or the MIT AI Risk Repository. "Attack" here means an incident coded as deliberate. This page has no list of records: those are on the pages it links to.

Incidents coded "Intentional" in the MIT causal taxonomy: the outcome was the expected result of someone, or something, pursuing a goal.

Synced 1,708 incidents on record newest dated 28 Sep 2026 classified to February 2026 No new record has arrived from the daily sync for 49 days; the figures stop there

In brief

735 of the 1,496 classified incidents in the AI Incident Database (49%) are coded as deliberate: the outcome was the expected result of pursuing a goal. 239 are dated in the 12 months to February 2026, up 12% on the 214 of the 12 months before. 128 newer incidents are recorded but not yet classified, so comparisons stop at February 2026.

Incidents in view
735
49% of the 1,496 classified
12 months to February 2026
239
+12% on the 12 months before (214)
Carried out by a person
63%
37% attributed to an AI system
Risk entries, same coding
537
from 74 frameworks, MIT AI Risk Repository
Not yet classified
128
incidents dated after February 2026

1. Rising: up 12% in the 12 months to February 2026

239 incidents coded as deliberate are dated in the 12 months to February 2026, against 214 in the 12 months before. The comparison stops at February 2026 on purpose: 128 newer incidents are on record without a classification, and counting those months would show a fall that is only a backlog.

Incidents per month, 36 monthsBy date of the incident. A month counts as classified once 75% of its incidents carry a code.
Incidents per month, 36 months 47 24 classified to Feb 2026 Oct 2023: 13 deliberate misuse of 18 recorded Oct 2023: 13 deliberate misuse of 18 recorded Nov 2023: 4 deliberate misuse of 10 recorded Nov 2023: 4 deliberate misuse of 10 recorded Dec 2023: 13 deliberate misuse of 22 recorded Dec 2023: 13 deliberate misuse of 22 recorded Jan 2024: 10 deliberate misuse of 17 recorded Jan 2024: 10 deliberate misuse of 17 recorded Jan 2024 Feb 2024: 16 deliberate misuse of 27 recorded Feb 2024: 16 deliberate misuse of 27 recorded Mar 2024: 17 deliberate misuse of 20 recorded Mar 2024: 17 deliberate misuse of 20 recorded Apr 2024: 17 deliberate misuse of 36 recorded Apr 2024: 17 deliberate misuse of 36 recorded May 2024: 12 deliberate misuse of 25 recorded May 2024: 12 deliberate misuse of 25 recorded Jun 2024: 11 deliberate misuse of 26 recorded Jun 2024: 11 deliberate misuse of 26 recorded Jul 2024: 13 deliberate misuse of 25 recorded Jul 2024: 13 deliberate misuse of 25 recorded Jul Aug 2024: 15 deliberate misuse of 24 recorded Aug 2024: 15 deliberate misuse of 24 recorded Sep 2024: 22 deliberate misuse of 28 recorded Sep 2024: 22 deliberate misuse of 28 recorded Oct 2024: 22 deliberate misuse of 34 recorded Oct 2024: 22 deliberate misuse of 34 recorded Nov 2024: 12 deliberate misuse of 18 recorded Nov 2024: 12 deliberate misuse of 18 recorded Dec 2024: 18 deliberate misuse of 26 recorded Dec 2024: 18 deliberate misuse of 26 recorded Jan 2025: 22 deliberate misuse of 32 recorded Jan 2025: 22 deliberate misuse of 32 recorded Jan 2025 Feb 2025: 33 deliberate misuse of 47 recorded Feb 2025: 33 deliberate misuse of 47 recorded Mar 2025: 30 deliberate misuse of 44 recorded Mar 2025: 30 deliberate misuse of 44 recorded Apr 2025: 27 deliberate misuse of 47 recorded Apr 2025: 27 deliberate misuse of 47 recorded May 2025: 22 deliberate misuse of 34 recorded May 2025: 22 deliberate misuse of 34 recorded Jun 2025: 28 deliberate misuse of 44 recorded Jun 2025: 28 deliberate misuse of 44 recorded Jul 2025: 19 deliberate misuse of 39 recorded Jul 2025: 19 deliberate misuse of 39 recorded Jul Aug 2025: 16 deliberate misuse of 35 recorded Aug 2025: 16 deliberate misuse of 35 recorded Sep 2025: 10 deliberate misuse of 17 recorded Sep 2025: 10 deliberate misuse of 17 recorded Oct 2025: 21 deliberate misuse of 41 recorded Oct 2025: 21 deliberate misuse of 41 recorded Nov 2025: 14 deliberate misuse of 28 recorded Nov 2025: 14 deliberate misuse of 28 recorded Dec 2025: 23 deliberate misuse of 44 recorded Dec 2025: 23 deliberate misuse of 44 recorded Jan 2026: 18 deliberate misuse of 38 recorded Jan 2026: 18 deliberate misuse of 38 recorded Jan 2026 Feb 2026: 11 deliberate misuse of 30 recorded Feb 2026: 11 deliberate misuse of 30 recorded Mar 2026: 11 deliberate misuse of 27 recorded (classification incomplete) Mar 2026: 11 deliberate misuse of 27 recorded (classification incomplete) Apr 2026: 8 deliberate misuse of 25 recorded (classification incomplete) Apr 2026: 8 deliberate misuse of 25 recorded (classification incomplete) May 2026: 6 deliberate misuse of 35 recorded (classification incomplete) May 2026: 6 deliberate misuse of 35 recorded (classification incomplete) Jun 2026: 0 deliberate misuse of 23 recorded (classification incomplete) Jul 2026: 0 deliberate misuse of 33 recorded (classification incomplete) Jul Aug 2026: 0 deliberate misuse of 13 recorded (classification incomplete) Sep 2026: 0 deliberate misuse of 15 recorded (classification incomplete)
  • Deliberate misuse
  • All recorded incidents
  • Classification incomplete
Incidents per month, 36 months
MonthDeliberate misuseAll recorded incidentsClassification
2023-101318complete
2023-11410complete
2023-121322complete
2024-011017complete
2024-021627complete
2024-031720complete
2024-041736complete
2024-051225complete
2024-061126complete
2024-071325complete
2024-081524complete
2024-092228complete
2024-102234complete
2024-111218complete
2024-121826complete
2025-012232complete
2025-023347complete
2025-033044complete
2025-042747complete
2025-052234complete
2025-062844complete
2025-071939complete
2025-081635complete
2025-091017complete
2025-102141complete
2025-111428complete
2025-122344complete
2026-011838complete
2026-021130complete
2026-031127incomplete
2026-04825incomplete
2026-05635incomplete
2026-06023incomplete
2026-07033incomplete
2026-08013incomplete
2026-09015incomplete

2. Who was responsible, was it meant, and what kind of harm

The widest path is a person acting with deliberate effect: 463 of the 735 classified incidents in view (63%). Follow a colour from left to right to see where deliberate and accidental incidents end up.

From cause to harm: responsible entity, intent and risk domainMIT causal and domain taxonomies, as applied by the AI Incident Database
From cause to harm: responsible entity, intent and risk domain RESPONSIBLE INTENT RISK DOMAIN Human → Deliberate: 463 (63%) AI → Deliberate: 271 (37%) Other → Deliberate: 1 (0%) Deliberate → Discrimination & Toxicity: 44 (6%) Deliberate → Privacy & Security: 53 (7%) Deliberate → Misinformation: 56 (8%) Deliberate → Malicious actors: 538 (73%) Deliberate → Human-Computer Interaction: 9 (1%) Deliberate → Socioeconomic & Environmental: 17 (2%) Deliberate → AI system safety, failures, & limitations: 18 (2%) Human 463 AI 271 Other 1 Deliberate 735 · 100% 1 Discrimination & Toxicity 44 2 Privacy & Security 53 3 Misinformation 56 4 Malicious actors 538 5 Human-Computer Interaction 9 6 Socioeconomic & Environmen... 17 7 AI system safety, failures... 18
  • Deliberate (coded Intentional)
From cause to harm: responsible entity, intent and risk domain
FromToIncidents
AIDeliberate271
HumanDeliberate463
OtherDeliberate1
DeliberateMalicious actors538
DeliberateMisinformation56
DeliberateHuman-Computer Interaction9
DeliberateDiscrimination & Toxicity44
DeliberatePrivacy & Security53
DeliberateSocioeconomic & Environmental17
DeliberateAI system safety, failures, & limitations18

3. Where incidents concentrate, year by year

The eight kinds of incident with the most records in view. A darker cell is a larger count. The last column is a part year, and its newest incidents are not classified yet.

Incidents by risk subdomain and year of the incident
Incidents by risk subdomain and year of the incident
Subdomain \ Year20192020202120222023202420252026
4.3 Fraud, scams, and targeted manipulation 3 7 4 12 42 100 183 37
4.1 Disinformation, surveillance, and influence at scale 2 6 2 6 21 49 37 8
3.1 False or misleading information — — 1 — 12 14 23 3
2.1 Compromise of privacy by obtaining, leaking or correctly inferring sensitive information 4 2 3 7 4 8 4 1
1.1 Unfair discrimination and misrepresentation 2 4 1 3 — 1 — —
1.2 Exposure to toxic content — 1 — 2 4 4 5 1
7.3 Lack of capability or robustness 1 1 2 2 2 2 1 —
4.2 Cyberattacks, weapon development or use, and mass harm — 1 — 1 1 1 5 —

4. What is described against what is recorded

The MIT AI Risk Repository codes 537 risk entries from 74 frameworks to the same subdomains as the incidents in view. Where the two shares differ, either the frameworks are ahead of the record or the record is ahead of the frameworks. Across 23 subdomains the two rankings agree weakly (Spearman's rank correlation 0.20).

Risk subdomains: share of risk entries against share of incidentsSelect a bubble to open the subdomain
Risk subdomains: share of risk entries against share of incidents 1% 1% 5% 5% 10% 10% 20% 20% 40% 40% 60% 60% Recorded more than described Described more than recorded Share of risk entries described by frameworks → Share of incidents recorded → 4.3 Fraud, scams, and targeted manipulation: 53.8% of incidents (395), 11.7% of risk entries (63) 4.3 4.1 Disinformation, surveillance, and influence at scale: 18.1% of incidents (133), 13.6% of risk entries (73) 4.1 3.1 False or misleading information: 7.4% of incidents (54), 0.9% of risk entries (5) 3.1 2.1 Compromise of privacy by obtaining, leaking or correctly inferring sensitive information: 6.1% of incidents (45), 1.9% of risk entries (10) 2.1 1.1 Unfair discrimination and misrepresentation: 3% of incidents (22), 0.4% of risk entries (2) 1.1 1.2 Exposure to toxic content: 2.5% of incidents (18), 1.3% of risk entries (7) 1.2 7.3 Lack of capability or robustness: 2% of incidents (15), 1.1% of risk entries (6) 4.2 Cyberattacks, weapon development or use, and mass harm: 1.2% of incidents (9), 12.7% of risk entries (68) 4.2 2.2 AI system security vulnerabilities and attacks: 1.1% of incidents (8), 15.5% of risk entries (83) 2.2 5.1 Overreliance and unsafe use: 1% of incidents (7), 1.3% of risk entries (7) 6.1 Power centralization and unfair distribution of benefits: 0.7% of incidents (5), 3.4% of risk entries (18) 6.2 Increased inequality and decline in employment quality: 0.7% of incidents (5), 3.5% of risk entries (19) 1.3 Unequal performance across groups: 0.5% of incidents (4), 0.2% of risk entries (1) 6.3 Economic and cultural devaluation of human effort: 0.4% of incidents (3), 2% of risk entries (11) 3.2 Pollution of information ecosystem and loss of consensus reality: 0.3% of incidents (2), 0.2% of risk entries (1) 5.2 Loss of human agency and autonomy: 0.3% of incidents (2), 1.1% of risk entries (6) 6.4 Competitive dynamics: 0.3% of incidents (2), 1.7% of risk entries (9) 7.4 Lack of transparency or interpretability: 0.3% of incidents (2), 0% of risk entries (0) 6.5 Governance failure: 0.1% of incidents (1), 0.4% of risk entries (2) 6.6 Environmental harm: 0.1% of incidents (1), 0.9% of risk entries (5) 7.1 AI pursuing its own goals in conflict with human goals or values: 0.1% of incidents (1), 9.5% of risk entries (51) 7.1 7.2 AI possessing dangerous capabilities: 0% of incidents (0), 9.7% of risk entries (52) 7.6 Multi-agent risks: 0% of incidents (0), 2.8% of risk entries (15)
  • 1 Discrimination & Toxicity
  • 2 Privacy & Security
  • 3 Misinformation
  • 4 Malicious actors
  • 5 Human-Computer Interaction
  • 6 Socioeconomic & Environmental
  • 7 AI system safety, failures, & limitations
  • Bubble size: incidents recorded. Dashed line: equal shares. Square-root scale. Unlabelled bubbles are named when pointed at.
Risk subdomains: share of risk entries against share of incidents
SubdomainShare of incidents (%)IncidentsShare of risk entries (%)Risk entries
4.3 Fraud, scams, and targeted manipulation53.839511.763
4.1 Disinformation, surveillance, and influence at scale18.113313.673
3.1 False or misleading information7.4540.95
2.1 Compromise of privacy by obtaining, leaking or correctly inferring sensitive information6.1451.910
1.1 Unfair discrimination and misrepresentation3220.42
1.2 Exposure to toxic content2.5181.37
7.3 Lack of capability or robustness2151.16
4.2 Cyberattacks, weapon development or use, and mass harm1.2912.768
2.2 AI system security vulnerabilities and attacks1.1815.583
5.1 Overreliance and unsafe use171.37
6.1 Power centralization and unfair distribution of benefits0.753.418
6.2 Increased inequality and decline in employment quality0.753.519
1.3 Unequal performance across groups0.540.21
6.3 Economic and cultural devaluation of human effort0.43211
3.2 Pollution of information ecosystem and loss of consensus reality0.320.21
5.2 Loss of human agency and autonomy0.321.16
6.4 Competitive dynamics0.321.79
7.4 Lack of transparency or interpretability0.3200
6.5 Governance failure0.110.42
6.6 Environmental harm0.110.95
7.1 AI pursuing its own goals in conflict with human goals or values0.119.551
7.2 AI possessing dangerous capabilities009.752
7.6 Multi-agent risks002.815
How incidents arise, against how frameworks expect them toPercent of the records that carry the code
  • Incidents recorded
  • Risk entries described

Responsible

A person
63%Incidents recorded
66%Risk entries described
An AI system
37%Incidents recorded
27%Risk entries described
Other or unclear
0%Incidents recorded
7%Risk entries described

Timing

After release
98%Incidents recorded
73%Risk entries described
Before release
2%Incidents recorded
8%Risk entries described
Other or unclear
0%Incidents recorded
19%Risk entries described

5. Who is harmed, and by what

The parties named most often as harmed, or nearly harmed, in the classified incidents in view, and the kind of risk behind each. Read a row as that party's risk profile and a column as the parties one kind of risk reaches. The record names values (epistemic integrity, privacy) as well as people, and a name is an allegation in a report, not a finding.

Who is harmed, by kind of riskCell: incidents. Shading: share of the row.
Who is harmed, by kind of risk. Each cell is the number of incidents; the shading is its share of the row.
Harmed party 1Discrimination & Toxicity 2Privacy & Security 3Misinformation 4Malicious actors 5Human-Computer Interaction 6Socioeconomic & Environmental 7AI system safety, failures, & limitations Incidents
Epistemic Integrity 3 2 22 174 · 1 · 202
General Public 6 6 25 144 · · · 181
Privacy 6 32 · 34 1 · · 73
Democratic Integrity 1 · 6 62 · · · 69
National Security And Intelligence Stakeholders · 1 5 43 · · · 49
Students 4 1 1 38 2 1 2 49
Women and girls 3 · · 35 · · · 38
Minors 2 4 · 29 · · · 35
  • 1 Discrimination & Toxicity
  • 2 Privacy & Security
  • 3 Misinformation
  • 4 Malicious actors
  • 5 Human-Computer Interaction
  • 6 Socioeconomic & Environmental
  • 7 AI system safety, failures, & limitations
Sector of deployment, where codedOnly 33 of 735 incidents carry a sector: a sample
Information and communication
13
Arts
7
Entertainment and recreation
7
Professional
4
Scientific and technical activities
4
Education
4
Financial and insurance activities
3
Administrative and support service activities
3

6. Where to look first

The questions a risk owner brings to this record, answered from the counts above. Each answer describes the record, not your exposure: use it to choose which subdomain to open.

Which kind of incident is recorded most?

4.3 Fraud, scams, and targeted manipulation: 155 of the 239 incidents coded as deliberate dated in the 12 months to February 2026 (65%).

Open 4.3: definition, frameworks and incidents

Which is growing fastest?

3.1 False or misleading information: 20 incidents in the 12 months to February 2026, against 15 in the 12 months before.

Open 3.1: definition, frameworks and incidents

What is recorded more often than the frameworks describe?

4.3 Fraud, scams, and targeted manipulation is 54% of the classified incidents here and 12% of the 537 matching risk entries in the MIT AI Risk Repository.

Open 4.3: definition, frameworks and incidents

What is described far more than it is recorded?

2.2 AI system security vulnerabilities and attacks: 16% of the risk entries and 1.1% of the incidents (8). A low count is not a low risk: a harm that has not happened yet, or is not reported, leaves no record.

Open 2.2: definition, frameworks and incidents

The board: one card for each kind of incident

Every kind with at least 10 records in view, busiest first. The label on each card is worked out from two counts, by the rules under the cards.

4.3 · Malicious actors

Steady

Fraud, scams, and targeted manipulation

2019 2026, part year
Year to Feb 2026
155
+16% (from 134)
All years
395
54% of view
Deliberate
100%
12% of risk entries

52 instruments on record for this risk domain

4.1 · Malicious actors

Steady

Disinformation, surveillance, and influence at scale

2019 2026, part year
Year to Feb 2026
41
-5% (from 43)
All years
133
18% of view
Deliberate
100%
14% of risk entries

52 instruments on record for this risk domain

3.1 · Misinformation

Rising

False or misleading information

2019 2026, part year
Year to Feb 2026
20
+33% (from 15)
All years
54
7.4% of view
Deliberate
100%
0.9% of risk entries

50 instruments on record for this risk domain

2.1 · Privacy & Security

Sparse

Compromise of privacy by obtaining, leaking or correctly inferring sensitive information

2019 2026, part year
Year to Feb 2026
4
-50% (from 8)
All years
45
6.1% of view
Deliberate
100%
1.9% of risk entries

16 instruments on record for this risk domain

1.2 · Discrimination & Toxicity

Sparse

Exposure to toxic content

2019 2026, part year
Year to Feb 2026
4
-20% (from 5)
All years
18
2.5% of view
Deliberate
100%
1.3% of risk entries

25 instruments on record for this risk domain

7.3 · AI system safety, failures, & limitations

Sparse

Lack of capability or robustness

2019 2026, part year
Year to Feb 2026
1
-50% (from 2)
All years
15
2% of view
Deliberate
100%
1.1% of risk entries

52 instruments on record for this risk domain

1.1 · Discrimination & Toxicity

Sparse

Unfair discrimination and misrepresentation

2019 2026, part year
Year to Feb 2026
0
-100% (from 1)
All years
22
3% of view
Deliberate
100%
0.4% of risk entries

25 instruments on record for this risk domain

Rising
at least 10 incidents in the last 12 months and up 25% or more on the 12 before.
Steady
at least 10 incidents in the last 12 months and within 25% of the 12 before.
Easing
at least 10 incidents in the 12 months before and down 25% or more since.
Emerging
fewer than 10 incidents in the last 12 months, at least three, and more than in the 12 before.
Sparse
too few recent incidents to state a direction.

7. What the rules on record cover

The same seven risk domains twice: the incidents in view, and the policy instruments in this tracker whose recorded use cases fall in that domain. An instrument can count under more than one domain. A long bar on the left beside a short one on the right is a question to ask, not a finding that the rules are missing.

Incidents in view, by risk domainClassified incidents only
1 Discrimination & Toxicity
44
2 Privacy & Security
53
3 Misinformation
56
4 Malicious actors
538
5 Human-Computer Interaction
9
6 Socioeconomic & Environmental
17
7 AI system safety, failures, & limitations
18
Instruments on record, by risk domainPublished instruments, by recorded use case
1 Discrimination & Toxicity
25
2 Privacy & Security
16
3 Misinformation
50
4 Malicious actors
52
5 Human-Computer Interaction
56
6 Socioeconomic & Environmental
142
7 AI system safety, failures, & limitations
52

How to read this page

  • These are reports, not a census. An incident is on record because someone wrote about it and an editor added it. Coverage leans toward English-language media, and a count is not a rate.
  • Classification trails the record. The codes for intent, entity and risk domain are added after an incident is recorded. A month is compared only when 75% of its incidents carry a code, which is why the year-on-year figures here end before the newest incident and differ from the calendar figure on AI risk.
  • "Deliberate" is a code, not a verdict. It is the MIT causal taxonomy's "Intentional": the outcome was the expected result of pursuing a goal. It covers fraud and impersonation as well as attacks on AI systems themselves.
  • Risk entries are counted the same way as the incidents in view. For "Deliberate misuse" only entries coded Intentional are counted; for "Security and malicious use" only entries in domains 2 and 4.
  • Updates. Incidents are synced from the AI Incident Database once a day and from its weekly export. Nothing on this page is fetched from the source while you read it.

Disclaimer: informational only, not legal advice. Verify every claim against the linked official sources and consult a qualified lawyer before acting.