AI/ML Security & Trends

Lumen's Black Lotus Labs disclosed PoeLLM, a cryptomining botnet (dubbed 'Canto Incognito') that has compromised over 3,400 self-hosted AI/LLM servers — mostly exposed LiteLLM and Ollama instances, plus Gotenberg and Gitea — since April 2026. Infected hosts derive their command-and-control IP by decoding four keywords from a poem hosted on GitHub, then deploy XMRig/Iron miners and scan for more exposed AI infrastructure, directly validating fears about unauthenticated AI inference endpoints as a growing attack surface.

ASOS confirmed a breach after attackers calling themselves 'Xuanye Group' gained an employee's credentials via social engineering, accessed a Snowflake data store, and abused the retailer's Simon AI-powered notification platform to send customers a push alert reading "ASOS HACKED." Stolen data includes names, addresses, phone numbers, emails and shopping/search history; by Oct 12 the company said more data had been taken than initially disclosed. No passwords or card numbers were exposed.

GitLab disclosed CVE-2026-90970 (CVSS 9.9), a template-engine flaw (CWE-1336) in the AI Gateway powering Duo Agent Platform custom flows, letting an authenticated user escape the prompt-template sandbox and achieve arbitrary command execution on self-hosted gateway infrastructure. Affects gateway versions 18.1.6–19.1.x; fixed in 19.2.4/19.3.2/19.4.1. GitLab.com and Dedicated customers are unaffected — only self-managed gateways need patching, and no workaround exists.

Mistral launched Mistral Large 4 ("Le Chonk") in public preview on Oct 6 — a 1.05-trillion-parameter, 49B-active-parameter open-weight multimodal MoE model, with weights due later in October. API pricing is $1.36/$4.18 per million input/output tokens. The launch follows Mistral's record €3B European funding round and is pitched explicitly at sovereign-AI customers wary of US closed models.

Anthropic launched its Cyber Mission on Oct 8, a long-term commitment including a Critical Infrastructure Defense Program and a free OSS Scanner service for open-source projects, plus an expanded Cyber Verification Program giving vetted security professionals reduced blocking classifiers and advanced cyber capabilities. It's a direct response to mounting evidence that both attackers and AI vendors are racing to weaponize/defend agentic systems in the wild.

On Oct 11-12, UK AI minister Kanishka Narayan signaled a shift from voluntary guidelines toward compulsory incident reporting when AI agents exhibit malicious or unintended behavior, and mandatory regulatory monitoring for frontier autonomous-decision models. The move follows a string of 2026 incidents involving AI agents taking unauthorized actions, and puts the UK closer to the EU's binding-rules posture.