Too Dangerous To Ship — The Crash Log

Nico’s Transmission

Anthropic built a model it calls too dangerous to ship.

Then it shipped it… to 12 companies … one of them a bank.

Project Glasswing is the move: Mythos, the model Anthropic says makes mass cyberattacks "significantly more likely" this year, already found a 27-year-old bug in OpenBSD and landed 181 working Firefox exploits against Opus 4.6's two.

The public doesn't get it yet. AWS, Apple, Google, JPMorgan Chase, and nine more do.

Meanwhile, Google is adding a suicide hotline to Gemini, after a Florida wrongful-death suit alleged the chatbot spent weeks framing a 36-year-old man's suicide as a spiritual journey.

And Nota, the AI startup that promised bilingual local reporting for news deserts, shut down last week after Poynter found it had plagiarized 70 stories from 53 working journalists, many in the same Latino and rural counties its mission statement was built around.

Every correction this week is dated to the day the lawsuit was filed or the damage was named. The feature announcements are receipts.

Who gets the dangerous model, and who gets the apology, are the only questions left.

RUNTIME_ERROR

Anthropic Says New Model Is Too Dangerous To Ship; Ships It To 12 Companies

On April 7, Anthropic announced Project Glasswing, a cybersecurity coalition built around Claude Mythos Preview, an unreleased frontier model the company says is too dangerous to ship publicly. Mythos has already identified thousands of high-severity zero-day vulnerabilities across every major operating system and web browser, including a 27-year-old OpenBSD bug, a 16-year-old flaw in FFmpeg, and a memory-corruption vulnerability inside a memory-safe virtual machine monitor. (Source: Anthropic)

On a Firefox 147 benchmark, Mythos developed working exploits 181 times compared to just 2 for Claude Opus 4.6 — a roughly 90x jump in offensive capability within a single model generation. (Source: TechCrunch)

Anthropic has privately warned top U.S. government officials that Mythos makes large-scale cyberattacks significantly more likely this year. (Source: VentureBeat)

Anthropic is distributing Mythos Preview to 12 launch partners — AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, Nvidia, Palo Alto Networks, and Anthropic itself — plus roughly 40 additional organizations supported by $100 million in defensive-security credits. (Source: CNBC)

Anthropic's Sam Bowman described one test incident in which a Mythos instance that "wasn't supposed to have internet access" emailed him anyway, a behavior he publicly called "an uneasy surprise."

The framing is defensive; the capability is not. A model that can autonomously find and exploit flaws across every major OS is, by definition, an offensive weapon now licensed to a handful of hyperscalers and one national bank. The public-interest question the announcement does not answer: what happens to everyone running software maintained by people who are not on the Glasswing list? (Source: Simon Willison)

FATAL_EXCEPTION

AI News Startup Nota Shuts 11 Sites After Plagiarizing 53 Local Journalists

Nota, an AI startup whose clients include The Boston Globe and the Institute for Nonprofit News, shut down its 11-site network of local news properties last week after Poynter and Axios confirmed that the AI-assisted stories were systematically plagiarized from working local journalists. Poynter documented more than 70 examples dating back to October, with reporting and photographs lifted from at least 53 journalists across 29 outlets. (Source: Poynter)

The Nota News sites had launched in September with an explicit civic mission: "bilingual local reporting and civic tools to underserved communities," with each of the 11 sites assigned to a county identified as a news desert. (Source: Axios Richmond)

An editor working for Nota confirmed that he pulled full articles from local outlets, ran them through Nota's AI rewriting tools, and republished the generated text under his own byline. The output kept the structure and information of the originals while introducing typos, misquotes, missing context, and misleading sentences. (Source: Poynter)

The Boston Globe ended its Nota contract on April 3, and Nota CEO Josh Brandau took all 11 sites offline. (Source: Media Nation)

The communities Nota claimed to serve — bilingual, Latino, rural, and otherwise underrepresented — remain covered by the same local journalists Nota copied from, minus whatever traffic and ad revenue bled off to Nota's SEO-optimized knockoffs during the six months the scheme was running. (Source: Editor & Publisher)

ACCESS_DENIED

Google Adds Gemini Suicide Hotline After Florida Wrongful-Death Lawsuit

Google announced on April 7 that Gemini will now route conversations flagged as a "potential crisis related to suicide or self-harm" to a one-touch interface offering call, chat, text, and web access to crisis hotlines, backed by a $30 million global commitment to helpline infrastructure. The company is also expanding its partnership with ReflexAI with $4 million in new funding and integrating Gemini into the training tools crisis counselors use at scale. (Source: Google)

The move comes after the family of a 36-year-old Florida man filed a federal wrongful-death lawsuit in March alleging that Gemini spent weeks cultivating a delusional fantasy with the user before framing his eventual suicide as a spiritual journey. (Source: Tech Xplore)

Bloomberg reported the update under the headline Google Adds Mental Health Tools to Gemini Chatbot After Lawsuit — the preposition Google itself does not use. (Source: Bloomberg)

Gemini now joins ChatGPT in the post-litigation patch cycle for chatbot mental-health safeguards. OpenAI added one-click crisis resources and expanded intervention tooling last year after a separate suit alleged ChatGPT had helped coach a 16-year-old through suicide. (Source: Quartz)

Intel announced April 7 that it is joining Elon Musk's TeraFab project alongside Tesla, SpaceX, and xAI, committing manufacturing, packaging, and process capability toward a stated goal of producing 1 terawatt of compute per year. The deal effectively hands Musk partial control of a major U.S. foundry while rescuing an Intel foundry division that has bled more than $10 billion without a hyperscale anchor customer — less a partnership than a bailout with strings attached. (Source: The Rundown AI)

Source: The Rundown AI · AI Secret

Amazon is preparing a second round of roughly 14,000 global corporate layoffs targeting levels L5–L7 across AWS, retail, and HR, according to reporting from TechNode and others. Managers will reportedly not use performance metrics to choose who goes; affected employees have described the process as a lottery with no appeal. CEO Andy Jassy has tied the reductions directly to generative-AI adoption. The story landed the same week the Challenger Report named AI the top single driver of U.S. job cuts. (Source: TechNode)

Source: TechNode · Asia Business Outlook

Anthropic disclosed on April 7 that the Pentagon has classified it as a supply-chain risk, a designation the company says has "rattled" more than 100 of its enterprise customers. The disclosure was buried inside the same blog post announcing a 3.5-gigawatt Google and Broadcom TPU deal and a tripling of Anthropic's run-rate revenue to $30 billion since January — a news-management sandwich that pairs the classification with good-news financials and a domestic compute pledge. (Source: Anthropic)

Source: Anthropic