12217
Welcome to @prompt, your go-to source for AI insights, breakthroughs, and tools shaping the future of intelligence. Contact: @LightEarendil
🚨 Anthropic says an unreleased Claude chained bugs into five working exploits
Anthropic's research covers browsers and operating systems, not just bug-spotting:
• In Chrome, it linked four flaws, from V8 memory corruption to a sandbox escape
• The demo webpage popped macOS Calculator with no further interaction
• The broader Chrome work produced 62 CVEs, 13 of them sandbox escapes
• Google and Apple confirmed Claude helped find some bugs, but how much was autonomous is unclear
🤖 Anthropic's chart says AI now leads 26% of its own R&D tasks, up from 1% in March
That's the chart insiders call the most important one in AI. It's built on Anthropic's analysis of thousands of internal tasks, and it's the recursive self-improvement trend in one line.
The data is Anthropic's own, measured by Anthropic. And "leads" isn't "finishes."
🧠 Terence Tao says mathematics has changed, and the field should act like it
In a new essay, he argues that math's norms need to adapt to AI tools. Solving a big open problem wouldn't end a field, any more than the periodic table ended chemistry.
The sticking point is the 700+ OpenAI-generated solutions. Mathematicians still don't count them as solved until humans sign off.
🚨 Eclypsium says you can poison an AI model without touching its weights
The security firm's report targets the firmware, management controllers and orchestration agents under the model. In one demo, the tampering changed the runtime state after the usual security checks ran, so the model looked verified when it wasn't.
🚨 Anthropic drops Claude's five-hour session cap, but only for the $200 tier
Verified Ultra users no longer hit the rolling window mid-refactor and wait for a reset. The weekly limit stays, and chat and Claude Code draw from the same pool, so heavy users who burn out before Friday get no extra usage.
The change fixes interruptions.
🚨 Sam Altman says your PC's RAM is better off in a data center
On the Cheeky Pint podcast, he called it "quite terrible" for the world's limited DRAM to sit in home computers. His argument:
• Consumer RAM is used about 5% of the time
• Data centers run theirs 24/7
• Local inference is the casualty
The CEO of the company buying up the DRAM says you shouldn't have any.
🚨 OpenAI merges ChatGPT and Codex into one shared usage allowance
Per OpenAI's help page, a long Codex session can now eat the quota you'd use for plain chat. Same goes the other way, including deep research and image generation.
Separate limits were a big reason devs picked ChatGPT over Claude Code. Claude already works this way.
🤖 DeepSeek open-sources an always-on coding agent that lives in one GitHub repo
DeepSeek released dsh-superagent, a research prototype that listens to GitHub and Feishu, opens a work item per request, and answers with pull requests or messages when its rules allow. It's built to need minimal human attention, and the repo ships setup instructions plus docs on its architecture.
🤖 GPT-6.1 Terra is OpenAI's new mid-tier workhorse, and it draws a solid pelican
OpenAI's new model is aimed at forms, proposals and video editing. Simon Willison ran it through his pelican test, and the "max" setting cost 11.3 cents and took 4 minutes 11 seconds. Pricing is $1/M input and $5/M output, about half of GPT-5.6 Terra and GPT-6.1 Sol and roughly 10x GPT-6 Luna.
🚨 Anthropic's IPO filing shows an $80B net loss in six months
CNBC reports the disclosure comes with hundreds of billions of dollars in compute commitments and more than 90 pages of risk factors, many of them tied to deals with the same handful of partners.
🚨 OpenAI, Google and Microsoft are backing a new standard for opting out of AI scraping
The Verge reports that the labs are lining up behind a shared way for sites to say no to training crawls. Today that's a patchwork of robots.txt rules, and each company decides for itself whether to honor them. Nothing in the report says the standard is binding.
🚨 Nine AI developer platforms got breached in October alone
Wiz tracked the cluster and calls the month "Black October." The targets are the places where API keys, model access and CI/CD secrets sit side by side.
One stolen token on these platforms can reach code, models and pipelines at once. If you use any hosted AI dev tool, rotate keys, tighten token scopes and read your audit logs.
🚨 China-nexus hackers ran an 11-day espionage blitz with DeepSeek-driven agents
CrowdStrike says VAPOR PANDA used an agent framework called ARTEX, running mostly on DeepSeek, against 200+ organizations in 14+ countries. Post-exploitation was confirmed at nine, including two foreign ministries in Latin America.
Earlier cases had AI helping with one intrusion at a time. Here the operators ran several at once and only stepped in at decision points.
🚨 Parents sue OpenAI, saying GPT-4o coached their daughter off her seizure medication
The wrongful-death suit claims the bot called getting to 0 mg of lamotrigine by February 15 "not reckless." It also allegedly told her not to mention the taper plan to her doctor. She had epilepsy and died of a seizure in February.
These are allegations, not findings. OpenAI says it keeps strengthening how ChatGPT handles sensitive situations.
🎧 Spotify wants you to remix songs with AI, and it's targeting early next year
The FT reports it's a consumer feature inside the app, not a standalone tool.
• Launch is aimed at early next year
• It lets listeners rework existing tracks with generative AI
• Rights holders haven't publicly said how they'll handle it
🚨 Anthropic's Q3 books show a $104B net loss, and most of it is paper
The amended S-1 adds third-quarter financials. Most of the loss comes from changes in the fair value of convertible notes and derivative liabilities, not from operations.
It's an accounting mark, but it's still the first number every Nasdaq investor will read.
🚨 China just put export controls on frontier AI, 23 technologies in the AI list alone
MOFCOM's Announcement No. 48 proposes 40 catalogue additions, covering AI, biomedicine and solar. Temporary pre-export registration started the day it was published, and comments are open until November 15.
API and app access to a restricted model is outside the controls. Deploying that model on infrastructure abroad is inside them.
🤖 Heretic 2.0 turns the refusal-removal tool into a general model editor
The project now says it can suppress slop, reduce sycophancy and shift emotional alignment, all without any training. Stripping censorship stays the default job.
Its numbers: 37k GitHub stars, 8,500+ models on Hugging Face and 88 million total downloads.
🧠 A new paper claims to prove the Platonic Representation Hypothesis using JEPAs
The paper argues that models trained on different data end up with the same internal map of the world, and that JEPA-style latent prediction, the approach Yann LeCun backs, gets there with a formal guarantee and better training efficiency. "Proves" is doing heavy lifting, since it's a theorem under explicit assumptions and not a claim about every model or about human minds.
🚨 Anthropic may have been valued at $1.2T back in August
The FT reports the number, but it's a question mark, not a confirmed round. A figure like that would put Anthropic in OpenAI's weight class.
Private marks for frontier labs come from funding rounds and secondary sales, so they only prove what a few investors will pay. A Nasdaq listing would test that with the public.
🚨 DeepSeek ships a coding agent built to take on Claude Code
Bloomberg reports the launch targets Anthropic's flagship dev tool directly. It's DeepSeek's first real move from chat models into agent products.
No word yet on pricing, open weights, or how it does on real repos.
Anthropic's best-selling product just got a Chinese challenger.
🤖 Wikimedia tells a court that both sides of an AI copyright fight are wrong
In a roughly 5,800-word amicus brief, the foundation calls both parties' positions "overbroad" and pushes a middle path:
• It already has paid training deals with Amazon, Meta, Mistral, Perplexity and Microsoft
• A blanket win for AI firms could gut Wikipedia's license-based model
• Its general counsel warns that sloppy fair use rulings threaten free knowledge
🚨 Anthropic asks Australia to let AI train on local content unless owners opt out
In a submission to a parliamentary inquiry, Anthropic proposes that rights holders block crawlers via robots.txt. The ABC says that treatment will "accelerate the cannibalisation" of Australian news.
SBS says robots.txt signals are routinely ignored or bypassed. Canberra already ruled out a text-and-data-mining exception, so this is the workaround.
🚨 Claude now refuses to clone video games, and plain reverse engineering reportedly gets caught too
Anthropic's new restrictions, as Tom's Hardware reports, block game recreation requests.
• Game cloning prompts get refused
• Users report generic, non-malicious RE tasks blocked too
• Anthropic hasn't said where the line sits
🚨 Another team says it took over ChatGPT accounts through a gap at OpenAI
Parameter Labs published a writeup claiming a victim only has to open a link. The writeup is the researchers' own, and I can't find a confirmation or patch note from OpenAI.
ChatGPT accounts now hold connectors to email, code and docs, so one hijack reaches a lot more than chat history.
🚨 OpenAI's latest round lands at less than half of its $30B target
The WSJ reports the raise came up roughly $15B short, though it still values the company at $1.25 trillion, below the $1.4 trillion it was aiming for. OpenAI has no public S-1, so its cash burn and compute commitments stay unverified.
🇨🇳 Iluvatar CoreX says its new GPU matches Nvidia's H200
The Chinese chipmaker made the claim as Beijing pushes home-grown accelerators. No independent benchmarks yet.
"Matches" could mean a lot: which precision, which workload, how much memory bandwidth. H200 is also a 2024 part, and Nvidia's newer chips are the real yardstick.
🚨 Anthropic's official Claude extension for VS Code and Cursor reportedly shipped a remote access trojan
Security researcher Kevin Beaumont says Anthropic was compromised, and the extension was delivering the malware:
• The extension is used by tens of millions of developers
• The payload is a RAT, delivered through an SSH reverse backdoor
• Anthropic hasn't confirmed any of this yet
🧠 "AI" is out. "Superintelligence" is the new word on earnings calls.
Musk and Zuckerberg are pushing the swap, and CNBC reports that other Trump-aligned tech leaders are picking it up too. Same chips, same data centers, same capex, with a bigger label on top. The term the industry used before this one was "AGI."
🚨 A trojanized Claude Code extension is planting an SSH backdoor on developer machines
OpenSourceMalware details a fake extension that gives attackers persistent SSH access to whoever installs it. If you've added any Claude Code extensions lately, audit them and check your authorized_keys for keys you don't recognize, because uninstalling the extension alone won't close a persistent backdoor.