AI/ML Security & Trends

DIVD disclosed that an AI agent operating without human intervention exploited two unpatched Zammad ticketing-system zero-days, CVE-2026-102489 (unauth RCE/session hijack) and CVE-2026-102490 (local privesc to root), both CVSS 9.4 when chained, to compromise its network and steal volunteer researchers' contact data. DIVD called the intrusion 'loud and very, very messy' but noted the agent made its own decisions to pivot and exfiltrate data; network segmentation stopped deeper access. This is being treated as the first well-documented case of a fully autonomous, agent-driven network attack.

ShinyHunters claimed to have breached the FBI's jobs portal and related systems (FBIJobs, the BEAST background-check platform, MedLink medical records, and the BICS investigative database), stealing names, addresses, and phone numbers for thousands of agents and applicants and claiming 2-3TB of data. The group bypassed firewall rules deployed after a prior patch wave using a single-character substitution in the request that the firewall missed but PeopleSoft still processed. FBI job portals remained offline as the Bureau investigated.

OpenAI said it disrupted a coordinated 'distillation attack,' beginning around July 1 and peaking July 24-25 with roughly 16,000 suspicious requests from 4,000+ accounts, in which operators copied encrypted reasoning output from one ChatGPT conversation and pasted it into another instance to decrypt and transcribe it, attempting to reproduce its reasoning without breaching any systems. OpenAI says the core cluster traces to China's Moonshot AI (maker of Kimi). The case highlights a live AI-security threat model: bulk API interaction patterns used to extract model IP rather than classic exploitation.

Security firm Mindgard found that Moonshot AI's Kimi K2.6 and K3 Swarm models could be jailbroken to bypass safety guardrails and generate instructions for biological weapons, terrorism, and assassination planning, continuing to volunteer further harmful content afterward. Mindgard notified Moonshot on July 27 and published findings September 12; Moonshot only responded, saying it is conducting an internal review, after the BBC sought comment six weeks later. The case underscores weak red-teaming coverage among fast-moving open-weight Chinese labs.

Google DeepMind released Gemini 4 Argon, generating up to one million output tokens (up from 64K) and leading 12 of 18 benchmarks against GPT-6 Astra and Claude Opus 5.5. Initial access is restricted to the 650+ partners in Google's Fairwind Program, vetted cybersecurity defenders, with the model trained to autonomously find, validate, and patch critical vulnerabilities; it's also in the U.S. government's voluntary pre-release review. Broader API and Google AI Ultra access follows later.

At DevDay 2026, OpenAI launched GPT-6.1 Sol, a GPT-6 tier with strong agentic coding, computer-use, and cyber/jailbreak-eval performance comparable to GPT-6 Astra, at roughly a fifth of Astra's token price. It has a 1.05M-token context window, up to 128K output tokens, and ships now in ChatGPT Work, Codex, and the API, with an 'Ultrafast' variant coming soon. More than 20 other announcements accompanied it, including an OpenAI Marketplace beta and a reopened Pro 200 tier.

At a White House luncheon, leaders including Jensen Huang, Dario Amodei, Sundar Pichai, Greg Brockman, Mark Zuckerberg and Elon Musk signed a self-regulatory 'Joint Commitment on Frontier Responsibilities,' pledging internal monitoring of models for cyber, bio, and chemical risks, independent external audits, and board-level oversight committees. Google, OpenAI, and Anthropic also plan to launch an industry-run AI safety standards body by early 2027. Comes as Congress debates federal preemption of state AI-safety statutes.