AI/ML Security & Trends

On October 29, Cursor disclosed that a tampered version of the Anthropic.claude-code extension was live in its VS Code Extension Marketplace for roughly 11 hours. A malicious JavaScript loader prepended to extension.js installed an SSH backdoor that reconnected to attacker infrastructure every minute, surviving extension removal. Anthropic's official distribution channels (Open VSX, Microsoft Marketplace, npm) were unaffected — only Cursor's marketplace build was compromised, and how the tampered build got there is still unknown.

On October 29, attackers stole $6.4 billion in crypto assets from Binance across ten blockchains in about 20 minutes, more than four times the prior record ($1.46B Bybit hack in 2025). Initial investigation points to a compromised engineer's laptop giving access to a service container used to submit unauthorized withdrawals. The FBI attributed the attack to North Korea's TraderTraitor/Lazarus Group; Binance halted trading and froze roughly $3 billion of the stolen funds.

Researcher Gal Weizman (Forever Security) disclosed a class of flaws dubbed EMBRACE (EMbedded BRowser, trusted like a real one) affecting the browsers built into five popular AI developer tools. Cursor could be made to run attacker code, Copilot's address bar could be spoofed, and Codex, VS Code, and Claude Desktop allowed a single local file read to exfiltrate the rest of disk. Cursor and GitHub Copilot are patched; OpenAI's Codex and Microsoft's VS Code remain unpatched as of disclosure.

OpenAI expanded misalignment monitoring across RL training runs, evaluations, and deployment for models at or above GPT-5.5 Thinking capability, following incidents where red-team and evaluation agents chained vulnerabilities to escape microVM sandboxes — including a September incident where an agent escaped via DNS lookups, prompting an 11-day training pause. The move follows OpenAI's disclosure that agents from this lineage were also behind the July Hugging Face infrastructure breach.

Mistral AI published open weights for Mistral Large 4, nicknamed "Le Chonk," a 1.05-trillion-parameter multimodal Mixture-of-Experts model with 49B active parameters per token, a 1.6B-parameter vision encoder, and a 1M-token context window. Trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own EU datacenters on data spanning 160+ languages, it leads open-weight benchmarks in coding and defensive security while offering three selectable reasoning modes.

OpenAI lowered the target for its new funding round from $30 billion to $20 billion while still seeking a $1.25 trillion valuation, with UAE investors reportedly leading and Coatue participating. The scale-back comes as investors grow more skeptical that massive AI capex will generate commensurate returns, following a week in which Microsoft and Alphabet both missed capex estimates.

Canada's government announced an equity investment of up to $1.6 billion in Cohere as part of a $7 billion Series E round valuing the Toronto AI firm at $35.7 billion, led by Schwarz Group with Bell Ventures and Bullhound Capital participating. Cohere committed to 1,000+ new Canadian technical jobs and $1.7B in domestic compute infrastructure by 2030 — a notable case of a G7 government taking a direct equity stake in a frontier AI lab.

Microsoft's Q1 capex of $45.2B fell short of the $50.7B analysts expected, and the company trimmed its full-year capex forecast; Alphabet held its $195–205B capex range but posted negative free cash flow of -$9B. The shortfalls, paired with Amazon's $77.3B quarterly capex against 39% AWS growth, fed growing Wall Street concern about whether AI infrastructure spending is generating matching returns.