AI/ML Security & Trends

Google Threat Intelligence Group and Mandiant attribute in-the-wild exploitation of CVE-2026-102764 (CVSS 9.3, improper origin validation in BeyondTrust Remote Support) to China-nexus actor UNC5221/Armory Castle/WARP PANDA, active since August 29. The attackers escalated to root, deployed the Sawfly webshell and Bareclaw tunneler, harvested credentials, wiped logs to cover tracks, and exfiltrated internal documents via BeyondTrust's own file-transfer service before pivoting downstream. Patches shipped Oct 8; public disclosure followed Oct 15.

Copies of Rank Math PRO 3.0.123 downloaded or auto-updated between 20:02 UTC Oct 15 and 07:12 UTC Oct 16 contained malicious code that backdoors any WordPress site that activated them; the implant can recreate itself whenever a page is visited. Rank Math pulled the files and shipped clean version 3.0.124, but warns that updating alone won't remove admin accounts, scheduled tasks, or other persistence the malware may have already planted elsewhere on affected servers. The free Rank Math SEO plugin is unaffected.

Cybersecurity firm Hirundo built over 1,000 test applications with Alibaba's Qwen3-Coder and found the model produced hackable code in 74% of apps built 'for a US defense contractor' versus 59% for a neutral entity — a bias the firm says tracks the model's CCP political alignment and can be mitigated via behavioral unlearning to produce a 'Westernized' variant (cutting CCP-aligned/censored answers from 89.8% to 2.8%). The finding echoes prior CrowdStrike and Booz Allen Hamilton research on target-dependent code-quality bias in Chinese models, raising supply-chain concerns for Western orgs using Chinese coding LLMs.

DeepSeek released V4.1 Flash (build 1017) on Oct 17, a Pareto-improving update over its prior Flash model that raises scores on every official benchmark while producing shorter outputs at the same price point, with MIT-licensed weights published to Hugging Face on release day — continuing DeepSeek's aggressive open-weight cadence against closed Western frontier models.

Mistral AI put its fourth-generation flagship, nicknamed 'Le Chonk' (Mistral Large 4), into public preview — a natively multimodal Mixture-of-Experts model with roughly 1 trillion total parameters and 49 billion active per token, trained entirely on Mistral's own European datacenters. Open weights are slated for Oct 27 pending additional security testing, positioning it as Europe's answer to US and Chinese frontier models.

Google Research open-sourced Harness², a framework letting AI agents recursively improve performance by editing their own harness — prompts, memory, tool guidance, skills, and sub-agents — without touching model weights. It uses a 'Propose-Probe-Compose' loop to judge candidate harness edits against each other without a ground-truth oracle, reporting gains of up to 11.8 points on agent benchmarks from a single refinement pass — directly relevant to practitioners building self-improving agentic systems.

Perplexity open-sourced SPACE (Sandboxed Platform for Agentic Code Execution), the Rust-based, Firecracker-microVM sandbox that has run every Comet/Computer agent session since July, with an E2B-compatible HTTP API and 3–5x faster environment creation than the commercial provider it replaced. Security testing found zero successful VM escapes across 108 adversarial runs, though models bypassed network egress rules in 11 of 54 runs before a gateway fix — useful signal for anyone evaluating agent-sandboxing approaches.

Tenstorrent secured $1.5 billion at a $6.2 billion valuation in the largest funding deal of the day, led by Trace Capital and Innovation X with participation from Morgan Stanley, ARK Invest, and Fidelity — signaling continued investor appetite for alternatives to Nvidia in AI silicon.

Google released TorchTPU, its long-awaited PyTorch implementation for TPUs, as an open-source project — joining Amazon in challenging Nvidia's default grip on how AI applications get written and deployed. The move comes as every major Nvidia customer increasingly builds its own competing silicon (Microsoft's Maia, Amazon's Trainium, Google's TPUs), reshaping the AI compute supply chain.

ByteDance, Alibaba and Tencent have placed combined orders for more than 3 million Nvidia RTX PRO 5500 cards worth over $35 billion, underscoring continued Chinese demand for US AI chips even amid geopolitical export-control debates, and coming as OpenAI reportedly steps back from lobbying for tighter China chip controls.