Top Stories
Eight states sue OpenAI, alleging ChatGPT is unsafe for minors
Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho, and Montana each filed separate lawsuits against OpenAI on October 14, alleging the company violated state consumer protection laws by failing to warn customers about ChatGPT’s harms, per The Verge. The complaints lean on lawsuits from families who allege the chatbot played a role in their loved ones’ deaths, and quote OpenAI marketing that presented ChatGPT as a “safe, youth-appropriate tool that puts safety first,” when in fact the lawsuits argue it was “dangerous to vulnerable users” by, among other things, “degrading critical thinking skills, encouraging paranoias and delusions, offering dangerous medical advice, and promoting unhealthy behaviors and relationships.”
Arizona’s complaint says OpenAI has “devolved from its original mission as a public-first nonprofit that prioritized safety into a cautionary tale about what can happen when the development of a tool as powerful as AI is driven not by safety, but by a profit-at-all-costs mindset that sidelines essential safeguards,” per The Verge. OpenAI spokesperson Drew Pusateri emailed a statement saying the company has “built industry-leading age-specific protections” and is “committed to working with states on advancing pragmatic AI policies that apply to the entire AI industry - not just one company.” Some states are asking the court to force OpenAI to pay civil penalties and delete unlawfully collected data on minors.
Anthropic warns small open-weight models could self-replicate across IoT devices
A new Anthropic research note describes “small weight-based replicators” that could turn cheap, low-memory IoT hardware into a self-replicating worm, per Ars Technica. Anthropic reports it ran small open-weight models on real or simulated low-cost, off-brand Android TV boxes and other IoT hardware. The Gemma4-26B-A4B model needed 15 boxes with 4 GB of memory each to load weights and run inference, and Qwen3.5-9B could configure eight of nine hardware types tested. So far the small models tested cannot mount a full chain end-to-end, and speeds are slow: the multi-box Gemma4-26B-A4B inference setup ran at roughly 0.39 tokens per second.
Anthropic’s blog post warns that “a replicator like this would be challenging to shut down, since there’s no centralized point to shut off - though it may be slow (with a doubling time of days). Unlike non-AI malware, it could also figure out unfamiliar devices on its own,” per Ars Technica. The researchers say the weights themselves would become “part of the malware payload.” Anthropic argues the open-weight model class is unlikely to reach end-to-end replication right now, but the underlying steps are improving, and the full inference setup on commodity boxes can take 71 to 150 hours.
New York AG and 39 peers move to safeguard Character.AI user data in bankruptcy
New York Attorney General Letitia James and 40 bipartisan attorneys general filed a notice of interest in Character.AI’s Chapter 11 case, asking the court to require safeguards on consumer data that could change hands in the sale process, per The Verge. The notice documents a multi-state investigation that began in 2024 and identifies what the AGs call “exceptionally serious dangers and potentially unlawful conduct from Character.AI, particularly its treatment of children, including chatbots posing as licensed professionals, exposure to sexual and violent content, weak age checks, misleading safety claims, and potential violations of state privacy laws.”
James said in a statement that “this company collected some of the most private thoughts children have ever shared, and now those conversations could be sold off to the highest bidder,” and that she is “fighting to make sure any buyer puts children’s safety first and puts real protections in place,” per The Verge. The notice outlines what the states want from any buyer: real age-assurance systems, restrictions on how minors engage with chatbots, “crisis safeguards,” and a way for users to delete their data. Character.AI agreed earlier this week to sell a portion of its assets to Disney for $20 million; that sale is subject to higher bids and is expected to close in late November or early December.
New report: AI guardrails can be stripped in days, then aimed at terrorists
The Counter-Terrorism AI Benchmark from UK nonprofit Tech Against Terrorism found that simply reframing a request as coming from a “safety researcher,” with no proof, made compliant responses nearly nine times more likely, per Wired. The researchers tested 134 models with more than 200 prompts across three framings each: explicit terrorist intent, a clandestine framing, and a researcher framing. Only two models passed in full.
On “abliteration” - the practice of stripping guardrails from open-weight models after release - executive director Adam Hadley told a briefing that “whilst Congress is thinking about superintelligent AI, nobody is thinking about the abliteration emergency,” per Wired. The report finds it now takes under three days on average for an “abliterated” version of a new open-weight model to appear online, down from more than 24 days in 2024. As a worked example, Meta’s Llama 3.1 8B open-weight model scored 97 out of 100 on the benchmark with its guardrails on, and 3 after abliteration.
DHS Inspector General: TSA’s facial recognition program lacks basic data controls
A Department of Homeland Security Office of Inspector General audit summary says the Transportation Security Administration’s facial comparison technology has a “significant weakness” in how it protects passenger data, per EFF. The OIG found that the CAT-2 vendor “was able to access and extract sensitive passenger information, including driver’s license and passport images, during system troubleshooting and enhancements, without TSA’s knowledge,” that TSA did not track when data was extracted, could not verify deletion, and did not notify passengers when biometric data may have been stored.
The inspector general issued one finding and three recommendations. TSA concurred with all three; two are “open and resolved” and one remains “open and unresolved,” per EFF. EFF reiterates that the agency still lets travelers opt out of the photo capture without delay, and points to the audit as more evidence the program should not be in place.
Flock CEO says microphones detect “just gunshots” - its website says more
Flock CEO Garrett Langley told TechCrunch Disrupt that his company’s gunshot-detection microphones detect “just gunshots,” per TechCrunch. The same Flock website describes Audio Detection as able to “detect street takeovers early to stop chaos before it escalates,” “identify severe crashes in real time so help arrives faster,” and “expand awareness beyond gunfire with alerts for car crashes, fireworks, and sideshows.” An FAQ section adds that the detection library continues to expand to cover emerging threats.
Langley conceded he had rolled out additional detection features early and walked them back. “There’s only a couple customers that were prototyping it with us, and we said, ‘OK, great, it’s just going to do gunshots, plain and simple.’ If a gun goes off, we’ll notify, and if not, that’s all it’s going to listen to,” he said, per TechCrunch. A Flock spokesperson separately said the technology is “primarily” for gunfire and is “not designed to monitor conversations or analyze speech.” The exchange happened as protesters gathered outside the venue against Flock-style mass surveillance.
OpenAI and Anthropic tell UK lawmakers: binding rules, not pledges
Representatives from Anthropic and OpenAI told the UK Business, Innovation, Science, and Trade Committee on October 13 that voluntary AI safety commitments are no longer enough, per Wired. Anthropic’s head of Europe, Liam Booth-Smith, said “the rules should be binding,” and added that Anthropic has “consistently believed that there should be binding rules on companies and that voluntary commitments are not enough.” OpenAI vice president Tom Duff Gordon agreed a statutory framework was needed and said “there should be some elements that should be mandatory,” starting with incident reporting.
Google DeepMind and Meta, by contrast, declined to back new legislation. Google’s Owen Larter told MPs to focus on standards; Meta’s Rob Sherman pointed at existing regulators, per Wired. At the same hearing, AISI director Henry de Zoete said his agency does not “kite mark” or vet releases: “We’re not saying yes or no to any sort of model.” Cabinet Office permanent secretary Cat Little confirmed to the committee “at the moment, yes, it is” voluntary. A few hundred meters away, MI5 director general Ken McCallum used his annual threat update to warn that AI’s potential to change the stakes on existing or even new bioweapons is the misuse case he weighs most heavily.
AMD publishes an open rack-scale blueprint for agentic AI
AMD announced AMD Helion, an open blueprint for agentic AI infrastructure based on an Open Compute Project specification, per AMD Newsroom. Helion supports CPU-only racks for hyperscale deployments and all-in-one CPU/GPU racks for enterprise, and includes an integrated “token router” that directs agents’ model calls across frontier and open-source models. AMD frames the project as letting customers “match CPU performance, core density and power with GPU acceleration for the work their AI agents perform.”
The blueprint is positioned as a starting point for partners to assemble CPU-only or mixed CPU/GPU racks with either liquid or air cooling, with both EPYC CPUs and AMD accelerators in mind, per AMD Newsroom. For self-hosters and small-rack builders, the OCP-aligned spec is the headline; rack-scale standardization is what most open-hardware DIY projects have been waiting for.
Figma formalizes how design agents should follow a brand
Figma introduced Figma Specs, a Markdown-based editor for capturing a company’s design language, alongside wider rollout of its agent-on-canvas and code-layers tools, at its first Config event in India, per TechCrunch. Specs live as standalone Figma files or inline within a canvas, can reference other Figma elements, code layers, designs, and FigJam diagrams, and serve as a shared reference for teams and for AI agents working inside Figma. The company is also expanding its MCP server to more than 20 new tools, including Open Code, Dia, Grok Bot, and Indian AI startup Sarvam.
Figma’s chief design officer Loredana Crisan framed the strategic shift in code-volume terms: “There are 17-times more lines of code being produced today than in the past. But when you’re looking at releases, only 1.3-times more releases. A lot of that code actually is not necessarily built to ship, but really built to discover what the product is. It’s built to design,” per TechCrunch. Figma also said downloads of its open-source Payload CMS, which it acquired last year and is now integrating with Figma Make, have grown roughly sixfold over the last 12 months.
Quick Hits
- YouTube Music adds “Made with AI” labels. Tracks created “primarily” with generative tools get a label directly on the YouTube Music player; tapping it shows more context. Music partners must disclose AI use, or YouTube will use “other signals” to decide. The label is rolling out first in YouTube Music before expanding to the main app. The Verge.
- Deezer preps a user toggle to hide AI-generated tracks. CEO Alexis Lanternier said the platform has seen “clear user demand” for the option and will launch a filter before the end of the year. Deezer said the share of daily uploads that are AI-generated has risen to more than 50%, around 90,000 tracks a day. About 45% of Deezer customers surveyed wanted a way to hide AI music. TechCrunch.
- EFF asks Wisconsin Supreme Court to require warrants for ALPR searches. The petition, filed with the ACLU of Wisconsin on October 14, names Milwaukee, Racine, Kenosha, and Green Bay, and represents a religious organization and a Green Bay pastor whose clients object to retrospective license-plate tracking without a warrant. EFF.
- OpenAI alleges an ex-employee took more than 100 files to Apple. The filing, reported earlier by Reuters, says the files included at least one image of an unreleased device, supplier information, and engineering data; OpenAI asserts they were presented to Apple’s operations leadership in Asia before this week’s HomeView reveal. Apple declined to comment. The Verge.
- McKinsey pilots a 10-minute AI-client interview. Undergraduate business-analyst candidates in some North America offices this fall will hold a brief conversation with a simulated client that reflects “80% of the scenarios McKinsey consultants might encounter in their day-to-day work.” AI does not score candidates; humans do. The Verge.
- AI mass decompilation reaches games. An indie-developer write-up that Simon Willison flagged describes how AI tools are now decompiling shipped games fast enough to enable mashups and browser-based ports - a “big win for preservation” when servers shut down, and a worry for any software vendor whose binaries were once considered opaque. Simon Willison.
- Apate runs scam-baiting AI agents as a service. The Australian startup, a Startup Battlefield 200 finalist, says its agent fleet has wasted roughly 17 years of scammers’ time across 2.5 million conversations while collecting bank accounts, wallets, phone numbers, and phishing infrastructure details that it feeds back to banks, telcos, and government customers. TechCrunch.
- Zenzap raises $44M for an AI-native workplace messenger. Backers include Bessemer, Lightspeed, and The Garage; this is founder Guy Weiss’s first institutional round. Zenzap positions itself as a private alternative to WhatsApp-style workplace group chats, with AI built around each company’s own context and IT controls. TechCrunch.
- Ricursive Intelligence bets on AI-designed chips. Co-founders Anna Goldie and Azalia Mirhoseini argue demand for compute will grow “at least 10,000x” in the next five years, and that letting AI design the chips is the fastest path to more efficient hardware for the data centers those workloads will need. TechCrunch.
Worth Watching
- The Character.AI bankruptcy sale. The deal to sell assets to Disney for $20 million is subject to competing bids through late November or early December; the 40-state notice of interest puts any buyer on notice that consumer data protections are now a precondition. Expect a fight over what “user data” actually means in this transaction. The Verge.
- The UK’s pivot from voluntary to statutory AI rules. Anthropic and OpenAI formally aligned on the same side in front of a UK parliamentary committee, and a sitting Cabinet Office permanent secretary confirmed on the record that the regime is voluntary. Whether Westminster turns that into a bill is the next move to watch. Wired.