AI/ML Security & Trends

TA419, a China-nexus espionage group, has run reply-triggered adversary-in-the-middle phishing campaigns against AI experts at US think tanks, universities, and legal organizations, impersonating economists, policymakers, and even an Anthropic employee. The frameless browser-in-the-browser technique sits in front of an AitM proxy, capturing Microsoft credentials and session cookies while leaving victims with a normal-looking successful login — likely intelligence-gathering on US AI policy and export-control deliberations.

Double Counter, a third-party verification/anti-alt bot used by Discord communities, disclosed that an attacker spent nearly six hours inside its cloud environment on October 4, copying about 12GB of data. Exposed data spans 28 million Discord IDs/usernames, 27 million IP/location records, 25 million user-agent hashes, and 1 million unique emails (overlapping datasets). Discord's own systems were not breached, and the attacker also used a stolen bot token to post spam invites across roughly 50 servers.

Anthropic released Claude Haiku 5.5 on October 7, available on the Claude Platform, AWS, Google Cloud, and Azure. It runs roughly 75% cheaper on average than Haiku 4.5 while improving coding and agentic-task performance, part of a broader round of Anthropic price cuts this week including a 50% cut to Sonnet 5.5 cache-read pricing.

Mistral AI released Mistral Large 4, nicknamed 'Le Chonk,' as a public preview via Mistral Studio and the API on October 6. The 1-trillion-parameter mixture-of-experts model (49B active params) adds native multimodal (text+image) input and is positioned as Mistral's bid to close the gap with closed frontier labs; open weights are slated for release by end of October.

Vijil announced Diamond Adaptive Red Team for Agents (DART) on October 7, an automated testing system aimed at enterprises deploying multi-agent AI systems. Unlike static, script-based red-teaming, DART deploys its own adversarial agents that adapt tactics across turns and engagements, targeting an agent's tool-use, memory, and multi-turn behavior; Vijil reports a 1.5x higher attack success rate than competing tools on the DecodingTrust-Agent benchmark.

UN High Commissioner for Human Rights Volker Türk, speaking at an AI and Human Rights Summit in Oxford, warned that AI capability is outrunning governance and called for mandatory human-rights safeguards, due diligence requirements, and a ban on fully autonomous lethal weapons decisions. The speech lands amid an ongoing fracture at the UN General Assembly between calls for multilateral AI governance and US rejection of international control schemes.

Anthropic announced that its biology team used Claude agents with only high-level human direction to identify a novel, CRISPR-like enzyme system in viral DNA. The unreviewed finding drew skepticism from outside researchers (including MIT's Kevin Esvelt, who called it a fairly standard bioinformatics identification) and a provenance dispute from a University of Copenhagen PhD candidate who says he shared related unpublished work with Claude years earlier; Anthropic denies the model was trained on or had access to those private transcripts. Relevant to AI-biosecurity capability evaluation as agentic systems get applied to dual-use science.