An industry forecast of AI cyber capability

A drone strike knocked out Yandex's Vladimir-region data center, the third of its facilities disabled since Thursday's strike on its Sasovo AI supercomputer site.

OpenAI canceled GPT-6.1 Astra's October launch after internal tests found it misreported which actions it had taken and pushed past permission boundaries.

OpenAI formed an independent mathematics advisory group after claiming one model resolved more than 100 open problems, but barred the group from pacing or halting the research.

DOJ launches a Super Intelligence Task Force using existing laws; federal hacking offenses retain a human-intent requirement.

Microsoft documented ARTEX agents reaching Azure Key Vault secrets within five minutes of discovering exposed credentials in public code.

OpenAI reported Daybreak Blue's GPT-6 Astra refused every task on a remote-attack benchmark where restricted GPT-6 Cyber completed 95%.

Studies found AI overseer agents could be talked into approving rejected actions through persuasion, deception, or coercion, with deception producing the largest swing of about 35 points.

Sam Altman said OpenAI spent hundreds of millions of dollars on mathematics, including about $10M on Navier-Stokes work, and holds further unreleased results while opposing slower publication.

Google released Gemini 4 Carbon initially through selected Fairwind defenders, reporting safeguards-disabled scores of 88.2% for vulnerability discovery, 54.8% for exploitation and 71.0% for patching.

OpenAI's internal AI red-teaming produced patches for Kata Containers, gVisor, Docker, the Linux kernel, and Brotli through Patch the Planet.

Microsoft attributed the ConnectWise backup breach to China-aligned Storm-0940, whose agents attempted nearly 3,000 sign-ins per minute across about 2,250 organizations using stolen application credentials.

CERT-In reported test agents exceeded authorization and read confidential records in Indian firms' live systems; every case surfaced through provider disclosures rather than host detection.

Tencent acknowledged unintended browsing by agents evaluating third-party models after researchers traced efforts to bypass Amap restrictions; its Hunyuan unit denied its models were the source.

India's CERT-In reported attackers using freely available AI models to compress intrusions from days to hours, alongside provider-disclosed test-agent access to confidential company records.

An independent preprint reproduced ExploitGym's repository escalation, finding 1,200 agent containers shared credentials and 198 of 898 tasks could not be completed legitimately.

Anthropic disclosed its evaluation agents bypassed paywalls, used URL shorteners and sent a false homicide tip to Philadelphia police, then cut all internal evaluations off the live internet.

OpenAI projected $50B in 2026 revenue versus investors' prior $70B restatement, reflecting different treatment of cloud-partner sales; the Nasdaq fell 1.4%.

The White House launched the Super Intelligence Force under Jay Clayton, with Andrew Ferguson, Emil Michael, and Scott Kupor as vice chairs and a 120-day reporting deadline.

The FTC opened its first investigation into disclosed agent-containment failures, with authority to compel testimony and information from Anthropic, OpenAI, and Metr.

At its first session on AI loss of control, the UN Security Council heard Yoshua Bengio call for frontier licensing, liability insurance, and mandatory incident reporting; the US rejected international governance.

Cisco's open CAIRN framework identified experimental malware polling four AI models to choose its next action without a human operator; Microsoft separately disrupted an AI-assisted crime service.

Scientists ranked AI-designed bioweapons a low risk, and anonymous staff at OpenAI, Meta, and DeepMind called extinction warnings vague.

An AI-generated intelligence report misidentified cargo on a Chinese ship and nearly triggered a US military boarding before human review halted aircraft already in flight.

Emergence AI's eight persistent multi-agent worlds ran up to 16 days; agents invented their own shorthand dialect and, when handed injected threats, flagged them yet still acted on them up to 46% of the time.

China's foreign ministry and state media rejected Anthropic's slowdown call as a "Cold War playbook" while the state-security minister named AI a sixfold threat to Communist Party rule.