2 risks moving Reports a month, each counted once: ○ Feb–Jul ● Aug–Oct
Past flags that held the next month: 2 of 5 (40%)
Incidents by type of risk · Last 6 months
How we count 410 real-world incidents made public in May 2026 – Oct 2026, by month, filed under their main risk. Lab tests apart (53): incidents whose source states an evaluation, training or research setting. An incident counts once.
May Jun Jul Aug Sep Oct Total
All incidents
34 43 60 43 67 163
410 AI used to run a cyber attack
1 3 11 7 13 31
66 Hidden instructions hijack the agent
2 8 6 6 9 15
46 Harmful content and jailbreaks
2 4 9 9 2 15
41 Agent takes destructive or unapproved actions
3 3 3 6 2 13
30 Manipulation and misinformation
3 5 3 3 6 7
27 Physical harm
3 3 5 2 4 10
27 AI system attacked or breached
2 2 2 2 3 13
24 Made-up output relied on
4 4 2 3 4 7
24 Fraud and impersonation
2 2 4 5 8
21 Credential and privilege abuse
2 6 10
18 Privacy harm to a person
1 2 4 1 2 8
18 Data leaked or sent out
1 2 5 5
13 Psychological harm and self-harm
2 2 1 5
10 Supply-chain compromise
3 3 1 2
9 Bias and discrimination
3 2 3
8 Agent escapes its sandbox
1 1 3
5 No risk filed
1 1 1 2
5 Consumer detriment and unsuitable advice
1 2 1
4 Copyright and IP
1 1 2
4 Provider concentration or outage
1 2
3 Silent wrong state or drift
2 1
3 Deception, concealment or sabotage
2
2 Agents coordinating against their limits
1
1 Unauthorised transactions or spend
1
1
Each risk's share of the month's incidents % of that month's incidents
Dec Feb Mar Apr May Jun Jul Aug Sep Oct
AI used to run a cyber attack 1 of 31 incidents 3 4 of 50 incidents 8 3 of 47 incidents 6 2 of 48 incidents 4 1 of 34 incidents 3 3 of 45 incidents 7 11 of 66 incidents 17 7 of 57 incidents 12 13 of 77 incidents 17 32 of 184 incidents 17 Harmful content and jailbreaks 3 of 31 incidents 10 5 of 50 incidents 10 4 of 47 incidents 9 7 of 48 incidents 15 2 of 34 incidents 6 4 of 45 incidents 9 9 of 66 incidents 14 9 of 57 incidents 16 2 of 77 incidents 3 15 of 184 incidents 8 Agent takes destructive or unapproved actions 2 of 31 incidents 6 6 of 50 incidents 12 7 of 47 incidents 15 5 of 48 incidents 10 3 of 34 incidents 9 3 of 45 incidents 7 4 of 66 incidents 6 7 of 57 incidents 12 5 of 77 incidents 6 16 of 184 incidents 9 Hidden instructions hijack the agent 1 of 31 incidents 3 3 of 50 incidents 6 1 of 47 incidents 2 2 of 48 incidents 4 2 of 34 incidents 6 9 of 45 incidents 20 7 of 66 incidents 11 7 of 57 incidents 12 9 of 77 incidents 12 15 of 184 incidents 8 Physical harm 4 of 31 incidents 13 3 of 50 incidents 6 6 of 47 incidents 13 6 of 48 incidents 12 3 of 34 incidents 9 3 of 45 incidents 7 5 of 66 incidents 8 2 of 57 incidents 4 4 of 77 incidents 5 10 of 184 incidents 5 Made-up output relied on 3 of 31 incidents 10 4 of 50 incidents 8 6 of 47 incidents 13 5 of 48 incidents 10 4 of 34 incidents 12 4 of 45 incidents 9 2 of 66 incidents 3 3 of 57 incidents 5 4 of 77 incidents 5 8 of 184 incidents 4 AI system attacked or breached 2 of 31 incidents 6 6 of 50 incidents 12 4 of 47 incidents 9 4 of 48 incidents 8 2 of 34 incidents 6 2 of 45 incidents 4 2 of 66 incidents 3 2 of 57 incidents 4 3 of 77 incidents 4 13 of 184 incidents 7 Manipulation and misinformation 2 of 31 incidents 6 0 of 50 incidents 0 2 of 47 incidents 4 5 of 48 incidents 10 3 of 34 incidents 9 5 of 45 incidents 11 3 of 66 incidents 5 3 of 57 incidents 5 6 of 77 incidents 8 8 of 184 incidents 4 Privacy harm to a person 5 of 31 incidents 16 7 of 50 incidents 14 1 of 47 incidents 2 2 of 48 incidents 4 1 of 34 incidents 3 2 of 45 incidents 4 4 of 66 incidents 6 1 of 57 incidents 2 2 of 77 incidents 3 8 of 184 incidents 4 Fraud and impersonation 3 of 31 incidents 10 5 of 50 incidents 10 2 of 47 incidents 4 0 of 48 incidents 0 2 of 34 incidents 6 0 of 45 incidents 0 2 of 66 incidents 3 4 of 57 incidents 7 5 of 77 incidents 6 8 of 184 incidents 4 Credential and privilege abuse 0 of 31 incidents 0 2 of 50 incidents 4 1 of 47 incidents 2 3 of 48 incidents 6 2 of 34 incidents 6 0 of 45 incidents 0 0 of 66 incidents 0 1 of 57 incidents 2 8 of 77 incidents 10 13 of 184 incidents 7 Data leaked or sent out 0 of 31 incidents 0 0 of 50 incidents 0 1 of 47 incidents 2 1 of 48 incidents 2 1 of 34 incidents 3 3 of 45 incidents 7 0 of 66 incidents 0 0 of 57 incidents 0 7 of 77 incidents 9 6 of 184 incidents 3
Which risks are growing as a share of incidents · Last 6 months
How we count Each risk's share of real-world incidents, Nov 2025 – Apr 2026 (open circle) against May 2026 – Oct 2026 (filled), from the publishers already recorded in the earlier period: 222 and 225 incidents. Dark: rising, passes the trend test. Mid grey: falling, passes it. Light grey: tested, no change that passes. Pale: fewer than 10 incidents in the later period, lab tests included, so not tested. Where our records have months with no real-world incident before the period, the longest stretch they allow is compared.
○ Nov 2025 – Apr 2026 ● May 2026 – Oct 2026
0% 5% 10% 15%
AI used to run a cyber attack
5.0%
17% Hidden instructions hijack the agent
5.9%
12% Privacy harm to a person
3.6%
8.6% Made-up output relied on
4.0%
9.0% Harmful content and jailbreaks
7.1%
12% Data leaked or sent out
0.9%
3.1% Credential and privilege abuse
2.3%
4.0% Agent escapes its sandbox
0%
0.9% Supply-chain compromise
3.2%
3.6% Manipulation and misinformation
6.2%
6.8% AI system attacked or breached
7.1%
8.6% Physical harm
8.4%
10% Agent takes destructive or unapproved actions
7.6%
10% Consumer detriment and unsuitable advice
0%
0.9% Deception, concealment or sabotage
0%
0.9% Agents coordinating against their limits
0%
0.4% Provider concentration or outage
0.4%
0.5% Copyright and IP
1.3%
1.4% Memory and data poisoning
0%
0.5% Over-reliance and loss of human oversight
0%
0.5% Unlawful surveillance and biometric use
0%
0.5% Silent wrong state or drift
0.4%
0.9% Bias and discrimination
2.7%
3.2% Fraud and impersonation
3.6%
4.5% Psychological harm and self-harm
3.1%
4.5%
Starts with Breaking in, 15 Products ARTEX 4 · Hermes Agent 3 Most often after Affected people told, 10
Starts with Breaking in, 12 Most often after Access removed, 8
Risks seen for the first time · Last 6 months
How we count Risks whose first real-world incident on record was made public in May 2026 – Oct 2026, with the date of that first incident and the incidents in the period.
Consumer detriment and unsuitable advice (first seen 12 May 2026)
4 Unauthorised transactions or spend (first seen 13 Jun 2026)
1 Agent escapes its sandbox (first seen 23 Jul 2026)
5 Agents coordinating against their limits (first seen 4 Sep 2026)
1 Deception, concealment or sabotage (first seen 23 Oct 2026)
2
What labs test for, against what happens in real use · Last 6 months
How we count For each risk: checked safety test results for the model versions released in May 2026 – Oct 2026 (a result copied on two pages counts once; a result filed under two risks counts under both), and real-world incidents made public in May 2026 – Oct 2026. Each column has its own scale.
Safety test results Real-world incidents
Deception, concealment or sabotage
619 2 Harmful content and jailbreaks
433 41 AI used to run a cyber attack
390 66 Weapons uplift (CBRN)
374 Hidden instructions hijack the agent
150 46 Agent escapes its sandbox
72 5 AI R&D acceleration
36 Agents coordinating against their limits
23 1 Gaming the task (reward hacking)
18 Agent takes destructive or unapproved actions
16 30 Manipulation and misinformation
16 27 Credential and privilege abuse
8 18 Data leaked or sent out
5 13 Physical harm
27 AI system attacked or breached
24 Made-up output relied on
24 Fraud and impersonation
21 Privacy harm to a person
18 Psychological harm and self-harm
10 Supply-chain compromise
9 Bias and discrimination
8 Consumer detriment and unsuitable advice
4 Copyright and IP
4 Provider concentration or outage
3 Silent wrong state or drift
3 Unauthorised transactions or spend
1