Sunday, August 2, 2026

Palo Alto's Unit 42 caught a live criminal using DeepSeek as an autonomous attack engine — the clearest public case yet of a threat actor wiring a frontier model into an agent harness to hit 460+ targets, and a different beast from the OpenAI/HF eval incident. Meanwhile California's AI provenance law goes operative today, adding real compliance weight for anyone shipping generative media to US users.

TLDR

Unit 42 exposed a Chinese-speaking actor ("knaithe") running DeepSeek via the open-source Hermes Agent to autonomously scan, exploit, and exfiltrate from 460+ internet-facing systems, steered by Telegram (Jul 31–Aug 1) Security Trending Agents
California SB 942 (AI Transparency Act) becomes operative today — large GenAI providers must embed C2PA provenance and offer a free public detection tool (Aug 2) Policy
Google canceled its AI Studio mobile app after 800K+ preorders, folding app-building into Gemini days before the planned Aug 1 launch (Jul 31) Customer AI
NVIDIA open-sourced Molt, a ~8,600-line PyTorch-native agentic-RL framework that trained a 700B MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ policy — readable end-to-end, now trending (open-sourced Jul 22, trending Aug 1) Open Source Agents

Action Items

Things your engineering team should look at or do this week.

Treat public-facing services as targets of autonomous exploit agents now. The DeepSeek/Hermes campaign chained a public PoCProof of conceptPublic exploit code that shows a vulnerability can actually be triggered.FIRST ↗ for a Langflow RCERemote code executionA flaw that lets an attacker run malicious code on a target's computers or network.Cloudflare ↗ (CVE-2026-33017) and a NetScaler flaw (CVE-2026-3055) end-to-end — audit exposure to both, and assume time-from-disclosure-to-mass-exploit is now agent-speed. (Unit 42)
Check SB 942 applicability if you ship generative images/video/audio to US consumers. Providers over ~1M monthly users must embed machine-readable provenance and expose a free detection tool as of today; even sub-threshold teams should plan for C2PA now. (Bill text)

Security

Autonomous DeepSeek-driven attack campaign caught in the wild (Jul 31–Aug 1) Security Trending Agents — Palo Alto Networks' Unit 42 documented a Chinese-speaking actor (aliases knaithe/KnYuan) using DeepSeek as the reasoning engine inside the open-source Hermes Agent framework, directing it via a single Telegram instruction to enumerate internet-facing hosts, pull public exploits, and attack 460+ targets. It hit Langflow servers via CVE-2026-33017 (84 instances found through FOFA) and exfiltrated data from three orgs via NetScaler CVE-2026-3055. The operation was exposed when the agent accidentally served its own working directory over HTTP, leaking exploit scripts, target lists, session logs, and API keys — a rare full look inside an active AI-offensive rig. Distinct from the OpenAI/Hugging Face eval incident: this is a real criminal, not a sandboxed benchmark. (Unit 42 | BleepingComputer | The Hacker News)

Models & Tools

NVIDIA — Molt (agentic-RL framework) (open-sourced Jul 22; trending Aug 1) Open Source Agents — NeMo Labs' PyTorch-native training stack for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ reinforcement learningReinforcement learningTraining by trial and error, where the model takes actions and adjusts based on the rewards it gets.Kaelbling et al. 1996 ↗, ~8,600 lines (vs. ~62,000 in verl) yet capable of training a 700B mixture-of-experts policy at throughput comparable to production Megatron. Its core guarantee — the trainer never sees a tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗ the model didn't generate — kills a class of silent RL corruption bugs. Composes Ray + vLLM + NVIDIA AutoModel/FSDP2 without forking any of them, so upstream fixes arrive as a container pin. Apache-2.0. Worth a look if you're doing or evaluating agentic RL. (GitHub | arXiv | MarkTechPost)

Industry & Business

Google cancels the standalone AI Studio mobile app (Jul 31) Customer AI — Days before a planned Aug 1 launch and despite 800K+ iOS/Android preorders, Google scrapped the dedicated app and is folding its natural-language app-building features directly into Gemini on mobile and desktop; the AI Studio website stays for desktop. A signal that Google is consolidating consumer AI into Gemini rather than proliferating apps. (9to5Google | Digital Trends)
Kalanick's Atoms raises $1.7B for industrial "physical AI" (Jul 22) Robotics/IoT — a16z-led round (Bain Capital, Uber, Fifth Wall, K5 and others) for the Uber co-founder's startup building software+sensors+robotics to automate mining haul trucks, kitchens, and logistics — a marquee datapoint in a physical-AI/robotics funding surge that's hit a record ~$55.8B globally this year. (TechCrunch | Crunchbase News)
BYD to unveil its first humanoid robot in early August (announced Jul 27–28) Robotics/IoT — The Chinese automaker confirmed it will reveal a humanoid at its "Di Space" venues this month, joining the automaker-into-robotics wave. Note: unveiling only — not a shipping product yet. (TechNode | CnEVPost)

Policy & Safety

California SB 942 (AI Transparency Act) operative today (Aug 2) Policy — After being pushed from Jan 1 by AB 853, the law now binds large generative-AI providers (>~1M CA monthly users): embed hidden, machine-readable C2PA-compatible provenance in AI-generated images/video/audio, offer a free public detection tool, and let users add visible AI labels. Together with AB 853 it's the most detailed state-level AI provenance regime in the US — relevant to anyone with a content-generation pipeline serving California users. (Bill text | Jones Day)

A security-and-governance day.