DeepSeek open-sourced a router-worker coding agent that cut task completion time nearly in half. DeepSeek released dsh-superagent under an MIT license, a research prototype that takes requests from GitHub and Feishu and distributes them across parallel worker sessions. In DeepSeek's own simulation, the 95th-percentile task completion time dropped from 29.5 minutes to 15.7 minutes. The router-worker architecture means a single incoming request can be split across multiple agents working simultaneously rather than sequentially. It is a research prototype, not a production product, so expect rough edges. (MarkTechPost)
Brief
Monday, November 2, 2026 · 4 min read
400 AI Apps Hid Deepfake Abuse Tools Behind Normal Interfaces
Good morning. Somewhere between the 400 apps hiding abuse tools behind a clean UI and OpenAI workshopping the phrase 'AI compute campus', it is a good day to think about what the word 'AI' is actually doing in a sentence. Sometimes it describes a capability. Sometimes it is a coat of paint over something that was already there, or already bad. Today has both.
Today's reading time is 4 minutes.
Listen · Snippt Daily0:00 / 4:07
SAFETY
App stores and the web have long hosted AI tools that look like photo editors or video generators, with no obvious way to know what else they do.
Driving the news: Wired identified roughly 400 apps and websites, linked to Chinese-language developers, that presented as ordinary AI tools but contained hidden features for generating violent and sexually abusive imagery. The tools were not niche dark-web products. They sat alongside legitimate apps, using the same design conventions, the same onboarding flows, the same category labels. The abusive features were accessible once a user was inside.
- The 400 figure covers both standalone apps and websites, not a single platform or developer.
- The hidden features included generation of violent AI imagery and non-consensual sexual deepfakes.
Zoom in: Non-consensual deepfake imagery has been illegal in a growing number of jurisdictions, but enforcement has focused on individual creators rather than the tools that make creation trivial. The app-store model assumes that surface-level review catches harmful functionality, and these tools exploited that assumption directly. A feature that only appears after account creation or inside a secondary menu is difficult to catch in a standard review process.
- The UK, several US states, and the EU have passed or are advancing laws targeting deepfake abuse, but platform liability for hosting the generation tools remains contested.
- Wired's count of 400 is a floor, not a ceiling - the investigation covered what researchers could find, not what exists.
Why it matters: For anyone building legitimate AI image or video tools, this is the context their product now sits inside. App store reviewers, enterprise buyers, and regulators are all going to be looking harder at what AI generation tools actually do once a user is past the front screen. The burden of demonstrating that a tool does not have hidden harmful modes is shifting toward developers, not away from them.
Bottom line: The threat model for AI-generated abuse imagery was never 'someone will build a dedicated abuse app' - it was always 'someone will hide it inside something that looks fine', and 400 examples just confirmed that.
Wired ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
NVIDIA released Auto Ontology, an open-source library that gives AI agents a shared vocabulary for querying enterprise data. Auto Ontology, released under Apache 2.0, generates a semantic layer from enterprise metadata - essentially a certified dictionary that tells an AI agent what a column or table actually means before it writes SQL. It serves that context through MCP (the protocol that connects agents to external tools), a CLI, and an API. The problem it addresses is real: agents querying databases without grounded definitions produce confident-sounding wrong answers. Whether an open library solves that at enterprise scale depends entirely on how well a company's metadata is maintained in the first place. (MarkTechPost)
OpenAI is reportedly considering calling its data centers 'AI compute campuses'. Data centers are unpopular neighbors - loud, water-intensive, and associated with industrial zoning fights. TechCrunch reports OpenAI is exploring the 'AI compute campus' label, which does not change what the facilities are but may change how they land in local planning conversations. It is a naming exercise, not an infrastructure announcement. The underlying tension, that AI's physical footprint is large and communities are noticing, does not go away with a rebrand. (TechCrunch)
The thread
Open weights, hidden features
Two of today's stories sit at opposite ends of the same transparency problem. DeepSeek's dsh-superagent ships MIT-licensed with its architecture documented and its benchmark numbers published - you can read what it does before you run it. The 400 apps Wired found did the opposite: clean surface, hidden functionality, no way to know from the outside. As open-source agent tooling gets easier to ship and harder to audit at scale, the gap between 'what the readme says' and 'what the binary does' is where the risk lives.
From the publisher
Snippt
Image, video and voice tools in one place
On our radar
- Nace AI open-sourced Drex Max v1, a 27-billion-parameter multimodal decision model that runs locally on Apple Silicon, scoring 67.52 on its benchmark - notable mainly because a model that size running on consumer hardware without a cloud call was not routine until recently.
- qwenGPU is a research project that compiles a 4-lane SIMT GPU into Qwen3.5-2B as 45 untrained layers, letting the model write and execute real CUDA launches itself - an unusual architecture experiment rather than a usable tool.
- Stories unrelated to AI tools: Heat director Michael Mann suffered minor injuries after a weapon discharged at a Los Angeles gun range.
- Also unrelated: Somali pirates seized an Iranian fishing vessel carrying 23 crew, with the EU's maritime security center warning it could become a base for further attacks on commercial shipping lanes.
AI worth acting on, in your inbox
Get AI Weekly every Friday: the AI news, tools and tactics that matter for marketers, sellers and creators. Two minutes, with audio.
You can change these choices or unsubscribe at any time. Privacy Policy