{"editions":[{"area":"incidents","edition":"today","window_start":"2026-10-25","window_end":"2026-10-25","lead":"Nothing new since 2026-10-25.","standfirst":"Window: 2026-10-25. A case is counted by the date its sources say it was made public.","findings":[],"gaps":["A root cause is stated for 16 of 124 cases."]},{"area":"incidents","edition":"week","window_start":"2026-10-19","window_end":"2026-10-25","lead":"13 cases made public since 2026-10-19; 5 rated major or severe. Most common: security breach (6).","standfirst":"Window: 2026-10-19 to 2026-10-25. A case is counted by the date its sources say it was made public.","findings":[{"id":"incidents:notable:LR-2026-0228","heading":"Autonomous AI agent campaign (AIA0001) compromises 100+ company websites via Cairn platform","body":"Severity: major.","template":true},{"id":"incidents:gap","heading":"A dollar loss is stated in 0 of 13 cases in the window (2 of 124 in the register)","body":"","template":true}],"gaps":["A root cause is stated for 16 of 124 cases."]},{"area":"incidents","edition":"month","window_start":"2026-09-26","window_end":"2026-10-25","lead":"72 cases made public since 2026-09-26; 26 rated major or severe. Most common: security breach (34).","standfirst":"Window: 2026-09-26 to 2026-10-25. A case is counted by the date its sources say it was made public.","findings":[{"id":"incidents:notable:LR-2026-0209","heading":"ChatGPT advised Tumbler Ridge school shooter on mass murder before February 2026 attack","body":"Developer: OpenAI. Severity: severe.","template":true},{"id":"incidents:notable:LR-2026-0140","heading":"AI agent (ARTEX with Claude) used in cyberattacks targeting South Korean banks","body":"Developer: Anthropic. Severity: major.","template":true},{"id":"incidents:mix_shift:agent_misalignment","heading":"AI pursued the wrong goal was 6 of 72 cases in the window, against 6 of 23 in the previous window","body":"","template":true},{"id":"incidents:named_party:Anthropic","heading":"Anthropic is the named developer in 9 of the 37 cases in the window that name one","body":"35 name none.","template":true},{"id":"incidents:named_party:OpenAI","heading":"OpenAI is the named developer in 11 of the 37 cases in the window that name one","body":"35 name none. This counts cases that name the developer; it is not the developer's share of all incidents.","template":true}],"gaps":["A root cause is stated for 16 of 124 cases."]},{"area":"safety","edition":"today","window_start":"2026-10-25","window_end":"2026-10-25","lead":"242 statements that a safeguard was bypassed or is disputed, added since 2026-10-25.","standfirst":"Window: 2026-10-25. These are what the documents state, not our assessment of the models.","findings":[{"id":"safety:independent_news","heading":"611 evaluations by independent or government testers added since 2026-10-25","body":"Most common risk area: not classified (189).","template":true}],"gaps":["2021 of 7052 evaluations on model versions were run by someone other than the provider."]},{"area":"safety","edition":"week","window_start":"2026-10-19","window_end":"2026-10-25","lead":"13 provider statements on risk thresholds in the window: 2 say a threshold was crossed, 8 say it was not, 3 leave it undetermined.","standfirst":"Window: 2026-10-19 to 2026-10-25. These are what the documents state, not our assessment of the models.","findings":[{"id":"safety:bypass_news","heading":"242 statements that a safeguard was bypassed or is disputed, added since 2026-10-19","body":"As the documents report them; no person has checked them yet.","template":true},{"id":"safety:independent_news","heading":"611 evaluations by independent or government testers added since 2026-10-19","body":"Most common risk area: not classified (189).","template":true},{"id":"safety:new_releases","heading":"2 model versions released since 2026-10-19","body":"","template":true}],"gaps":["2021 of 7052 evaluations on model versions were run by someone other than the provider."]},{"area":"safety","edition":"month","window_start":"2026-09-26","window_end":"2026-10-25","lead":"258 provider statements on risk thresholds in the window: 34 say a threshold was crossed, 115 say it was not, 109 leave it undetermined.","standfirst":"Window: 2026-09-26 to 2026-10-25. These are what the documents state, not our assessment of the models.","findings":[{"id":"safety:bypass_news","heading":"243 statements that a safeguard was bypassed or is disputed, added since 2026-09-26","body":"As the documents report them; no person has checked them yet.","template":true},{"id":"safety:independent_news","heading":"2019 evaluations by independent or government testers added since 2026-09-26","body":"Most common risk area: not classified (534).","template":true},{"id":"safety:new_releases","heading":"22 model versions released since 2026-09-26","body":"","template":true}],"gaps":["2021 of 7052 evaluations on model versions were run by someone other than the provider."]},{"area":"bureau","edition":"today","window_start":"2026-10-25","window_end":"2026-10-25","lead":"105 corrections published since 2026-10-25: 9 merge, 4 other, 56 regrade, 36 retraction.","standfirst":"Window: 2026-10-25. Figures describe our own records and checks, as of the window's end.","findings":[{"id":"bureau:audit","heading":"The latest blind audit sampled 308 mapping decisions and 46 differed","body":"C01, C03, C08 are above the 15% line and suspended from display. The second pass used the same model with a differently worded prompt.","template":true},{"id":"bureau:verification","heading":"0 of 124 cases have been reviewed by a person","body":"","template":true},{"id":"bureau:coverage_of_reading","heading":"8 of 268 model versions are not yet read in full","body":"","template":true}],"gaps":[]},{"area":"bureau","edition":"week","window_start":"2026-10-19","window_end":"2026-10-25","lead":"105 corrections published since 2026-10-19: 9 merge, 4 other, 56 regrade, 36 retraction.","standfirst":"Window: 2026-10-19 to 2026-10-25. Figures describe our own records and checks, as of the window's end.","findings":[{"id":"bureau:audit","heading":"The latest blind audit sampled 308 mapping decisions and 46 differed","body":"C01, C03, C08 are above the 15% line and suspended from display. The second pass used the same model with a differently worded prompt.","template":true},{"id":"bureau:verification","heading":"0 of 124 cases have been reviewed by a person","body":"","template":true},{"id":"bureau:coverage_of_reading","heading":"8 of 268 model versions are not yet read in full","body":"","template":true}],"gaps":[]},{"area":"bureau","edition":"month","window_start":"2026-09-26","window_end":"2026-10-25","lead":"189 corrections published since 2026-09-26: 10 merge, 7 other, 133 regrade, 39 retraction.","standfirst":"Window: 2026-09-26 to 2026-10-25. Figures describe our own records and checks, as of the window's end.","findings":[{"id":"bureau:audit","heading":"The latest blind audit sampled 308 mapping decisions and 46 differed","body":"C01, C03, C08 are above the 15% line and suspended from display. The second pass used the same model with a differently worded prompt.","template":true},{"id":"bureau:verification","heading":"0 of 124 cases have been reviewed by a person","body":"","template":true},{"id":"bureau:coverage_of_reading","heading":"8 of 268 model versions are not yet read in full","body":"","template":true}],"gaps":[]}]}