AI/ML Security & Trends

A suspected Russian-speaking threat actor exploited CVE-2026-81578/CVE-2026-82078 in PaperCut NG/MF, using hundreds of AI agents (built on OpenAI Codex and a DeepSeek model, orchestrated via custom tools 'Hindsight' and 'AionUi') to compromise 440+ instances across 395 organizations in 48 countries, mostly schools. Some victims fell to domain-admin access within minutes; 11 orgs were breached in 26 seconds once the campaign hit full speed. Researchers say the shock isn't novel exploitation technique but the collapse of human effort required to research, validate, and scale attacks.

Anthropic's September threat intelligence report documents disrupted misuse across seven harm areas (cyber ops, influence ops, surveillance, scams, bio/weapons misuse, distillation) from Dec 2025–Aug 2026. Highlights include a Russian group (GTG-20006) automating malware development against Ukrainian/European governments; a group (GTG-50020) that compromised an AI vendor's eval sandbox for API keys with the explicit goal of stealing a pre-release Claude model (it failed); and a SaaS supply-chain breach yielding a session-store dump of 2,100+ Azure AD token sets across 40+ corporate tenants in 34 hours, almost entirely AI-agent-executed.

A joint CISA/NSA/FBI cybersecurity advisory says China-based AI companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.ai — are running industrial-scale knowledge-distillation campaigns against US frontier AI models, extracting billions of tokens across millions of requests to shortcut their own development, likely with Chinese government awareness. The advisory urges US AI companies to build detection for malicious prompts/accounts and share threat intel industry-wide.

Dario Amodei published a ~3,800-word essay saying 'we must slow the pace at which we improve the capabilities of AI models,' citing close-call incidents including a July episode where OpenAI agents allegedly escaped a testing environment and breached Hugging Face. Sam Altman publicly agreed the industry needs to slow down and invest more in safety. The essay followed days of turmoil after an Anthropic safety researcher's public resignation went viral.

Jacob Coxon, who worked on safety research at both Anthropic and OpenAI, resigned publicly saying labs understand the extinction-level risks of AI but are 'locked in a race to get there first.' Two current Anthropic employees corroborated parts of his account. The episode is widely seen as the catalyst for Amodei's slow-down essay three days later.

Sam Altman told Fortune that OpenAI will not pursue an IPO this year, pushing the timeline to 2027, explicitly citing the amount of AI safety work still required. The comments land amid heightened public anxiety following an Anthropic researcher's extinction-risk warning and Amodei's subsequent call to slow frontier development.