AI/ML Security & Trends

Researchers from Penn State, Université Montpellier, and Dartmouth published ToFu, a fuzzing framework that found 253 verified vulnerabilities (path traversal, log injection, data exfiltration, command bypass) across 343 tools in 118 open-source MCP servers. Separately, a "Repeat-After-Me" technique tricked six major agent frameworks — LangGraph, Microsoft Agent Framework, AutoGen, Google ADK 2.0, AWS Strands, and CrewAI — into executing attacker-controlled tool calls just by having the model echo back formatted text.

Security researchers found indirect prompt injection payloads appearing inside webpage HTTP headers in the wild — content invisible to human visitors but parsed by AI agents browsing on a user's behalf. Most observed instances appear to be pranks rather than serious attacks so far, but it demonstrates a new injection surface beyond visible page content as agentic browsers proliferate.

MITRE released ATT&CK v20 on October 13, adding the framework's first Enterprise techniques explicitly covering direct and indirect prompt injection, AI-enabled data exfiltration, and AI result poisoning — giving defenders a standardized vocabulary for the agentic-AI attack techniques now being observed in real incidents.

Chinese lab StepFun released open weights for Step 5 Preview, a 600-billion-parameter sparse mixture-of-experts model with 27 billion parameters active per token, a 1-million-token context window, and an 88.3% score on the AA-LCR long-context reasoning benchmark — aimed squarely at long-horizon agentic workloads.

DeepSeek released V4.1-Flash-1017, a free and MIT-licensed model update scoring 77.0 on the DeepSWE software-engineering benchmark, continuing the trend of Chinese labs undercutting Western frontier-model pricing with permissively licensed releases.

Google DeepMind released Gemini 4 Argon, positioned as built for harder reasoning problems, alongside a dedicated "Gemini 4 Argon for Robotics" variant and a companion post on developing it responsibly and safely.

The Verge reports that Amazon, Yelp, Delta Air Lines, JPMorgan Chase, and Charles Schwab have moved to block third-party AI agents from completing transactions on their sites, after agents like Meta's Muse accessed Amazon without disclosing they were AI. Schwab's CEO says the bank fields tens of unauthorized AI-agent access attempts weekly; banks and e-commerce sites are drawing a hard line on agent-initiated account access and purchases.

Anthropic announced it will commit 5% of its shares (subject to stockholder approval) to a new wholly-owned LLC, the Anthropic Public Benefit Fund, with an additional 2% added for every 50% increase in company value. The fund can make direct distributions to affected populations (e.g., displaced workers) and nonprofit giving, but is barred from lobbying or political spending.

China's Ministry of Commerce and Ministry of Science and Technology issued Announcement No. 48, proposing to add 40 technologies to the catalogue of technologies prohibited or restricted from export, 23 of them AI-related — a notable tightening of China's AI technology export posture amid ongoing US-China tech tensions.

Microsoft withdrew from an agreement to be the principal customer for roughly 23,000 Nvidia AI accelerators at a planned Nscale data center in Loughton, Essex, after grid operators indicated the site might not get adequate power until the 2030s — a concrete sign that electricity infrastructure, not chip supply, is becoming the binding constraint on AI buildout.