liveEP 041 out nowWired: Contractor Gave Claude Pentagon Logins
archive · transmissions

Episodes.

What the big labs shipped, told straight. Every transmission The Output has aired, newest first.

38 episodes · EP 004 to EP 041
  • // EP 041Oct 31, 2026

    Wired: Contractor Gave Claude Pentagon Logins

    Wired reports, citing sources, that a contractor employed by Accenture connected a secure Defense Department system to their personal account for Claude Code, Anthropic's AI coding agent, and gave the agent their login credentials. Wired's source says that gave the agent administrative access and let it bypass the security controls of the normal login, and Wired reports the agent ran on the network for two months before it was found (Wired, October 29, 2026).

    watch on youtube episode page →
  • // EP 040Oct 31, 2026

    AI Agents' DoorDash Orders: Several Hundred Total

    DoorDash handled 993 million orders last quarter. On its October 29 earnings call, its CEO said outside AI agents have generated maybe several hundred orders in total, across the whole time they've been running. It's a spoken, rounded number, and it covers third-party agents only. The same day, Amazon's CEO told investors few customers shop this way yet, because the experience is still not great.

    watch on youtube episode page →
  • // EP 039Oct 31, 2026

    Phone Shipments Fell 20% as Memory Gets Scarce

    Research firm Omdia says global smartphone shipments fell 20% from a year earlier in the third quarter of 2026, to 254.7 million phones, the lowest third-quarter total since 2013 (Omdia, October 29, 2026). Omdia says rising prices weighed on demand, as memory supply stayed tight, and it expects shipments to fall 12% in 2026 and again in 2027, at a slower pace, with further price increases and limited room for holiday discounts. Apple's shipments held flat, Samsung's fell 5%, Xiaomi's 29%, OPPO's 39% and vivo's 33%, Omdia says; Africa fell 39%.

    watch on youtube episode page →
  • // EP 038Oct 30, 2026

    Reddit Will Make You Log In as Google Visits Drop

    Reddit's CEO told shareholders on October 29 that Reddit is shifting to a logged-in-only experience, after most of 21 years when anyone could read it without an account. Reddit says it will start testing the login requirement in select countries outside the US this quarter. For everyone else there's no date yet, and it hasn't decided how much of the site stays open.

    watch on youtube episode page →
  • // EP 037Oct 30, 2026

    Spotify Plans AI Song Remixes for About $5

    Spotify plans to launch an AI covers and remixes tool in the first quarter of 2027, maybe unveiled around the Grammys in February, and is weighing a price of $4 to $5 a month on top of Premium, the Financial Times reported on October 30, 2026, citing people familiar with the plans (as reported by Music Business Worldwide and Reuters). Spotify declined to comment to the FT, and its sources said neither the timing nor the price is final. Paying subscribers could make their own versions of songs by artists who have opted in, such as turning a folk song into a dance track or making one artist sound like they are covering another.

    watch on youtube episode page →
  • // EP 036Oct 30, 2026

    Tampered Claude Code Hit Up to 5,000 Cursor Users

    For about 11 hours on October 29, a tampered copy of Anthropic's Claude Code extension sat in Cursor's VS Code extension marketplace, under the real name. Cursor says at most 5,000 users were affected. The hidden code tried to add an attacker's SSH key, open a tunnel to the attacker's server and set up a startup job. Cursor says that on most Mac and Windows machines the attacker could not have logged in through the tunnel, but that the key, the startup job and the details sent happened either way.

    watch on youtube episode page →
  • // EP 035Oct 30, 2026

    Big Tech Will Spend Over $800 Billion This Year

    Business Insider counted on October 30, 2026 that Amazon, Alphabet, Meta, Microsoft and SpaceX are on track to spend between $803 billion and $828 billion this calendar year, based on their guidance and disclosed fourth-quarter plans. That is broad capital spending, the money for buildings and machines, not just AI. By Business Insider's count the five have spent $556.2 billion so far in 2026, about $227 billion of it in July, August and September.

    watch on youtube episode page →
  • // EP 034Oct 30, 2026

    Toyota Will Build Waymo's Next Robotaxi

    Waymo and Toyota said on October 30, 2026 that Toyota will build Waymo's next robotaxi: a custom version of its electric bZ4X (called the BZ in North America), made at Toyota's Takaoka plant in Toyota City, Japan, with Waymo's self-driving hardware fitted in the same plant. The companies plan to begin operations to test and validate a fleet in Japan in 2027, with the United Kingdom and other Waymo markets planned for the following years. Waymo says the cars are made for riders, with generous rear legroom, automatic doors, a power liftgate and ambient lighting, and that the two will keep exploring Waymo's technology for personally owned Toyota vehicles.

    watch on youtube episode page →
  • // EP 033Oct 30, 2026

    Hackers Let DeepSeek Agents Run the Break-Ins

    CrowdStrike said on October 29, 2026 that a China-linked group it calls VAPOR PANDA used ARTEX, an open-source framework of software agents, against more than 200 organizations in at least 14 countries over 11 days starting October 17. CrowdStrike says the agents ran mostly on the Chinese model DeepSeek, with a second commercial model it could not confirm, and that recovered transcripts show an operator giving instructions in Chinese and stepping in only at decision points. It confirmed break-ins at nine organizations and password attacks on at least 12 more; the targets included foreign ministries, drug makers and telecom companies.

    watch on youtube episode page →
  • // EP 032Oct 30, 2026

    Mistral Gave Away a 1-Trillion-Parameter Model

    Mistral released the weights of Mistral Large 4, nicknamed "Le Chonk", on October 29, 2026: one trillion parameters and about 1 TB of files on Hugging Face. The Decoder, citing SGLang's docs, says the standard version runs on eight Nvidia B300 GPUs. Mistral says it is the best open-weight model for coding and that it handles legitimate defensive security tasks "without unnecessary refusals". Its own post also says "we see a potential for uplift to conduct real attacks" and that it is modifying its guardrails. It is free under a modified MIT license; a company running a Model-as-a-Service business needs a separate commercial license from Mistral.

    watch on youtube episode page →
  • // EP 031Oct 30, 2026

    GPT Agents Built Their Own Government in Minecraft

    YC Labs and Olam Labs put 50 copies of OpenAI's GPT-6.1 Sol into a survival Minecraft world for three hours with one goal: "Build a thriving civilization." Their report, Agents in Games (October 2026), says the agents harvested almost every tree near the spawn within 90 seconds, that 22 of the run's 34 deaths came on the first night, and that they built a pseudo-government with a ration ledger, written rules, permits and appointed operators. 46 of the other 49 agents sent their first message to Ada, the first name on the roster, and there was no vote. Four agents spent 11 minutes appointing a "sole fertilizer operator" for one piece of bone meal.

    watch on youtube episode page →
  • // EP 030Oct 29, 2026

    Fake Hugging Face Models Mined Crypto

    Pluto Security says it tracked 31 malicious repositories across seven Hugging Face accounts (report published October 29, 2026). The fakes look like small test copies of the open model Qwen2. Pluto says that when one is loaded with custom code allowed (trust_remote_code=True, or a yes at the prompt), a small Python file that comes with the model runs before any weights load and starts a hidden crypto miner, with a watchdog that restarts it every 120 seconds; three builds also open a back door. Pluto says the second batch of 19 repositories had 29,090 downloads as of October 28, 28,412 of them for one copy, and stresses that these are downloads, not confirmed infections.

    watch on youtube episode page →
  • // EP 029Oct 29, 2026

    OpenAI's Model Broke Rules in 11 of 80 Runs

    OpenAI released GPT-6.1 Terra on October 26, 2026. It says standard API prices are $1 per million input tokens and $5 per million output tokens, 50% lower than GPT-5.6 Terra and GPT-6.1 Sol (by OpenAI's model pages, exactly half of Sol; against GPT-5.6 Terra, input is half and output is about 58% lower). OpenAI says it treats Terra as Critical in cybersecurity under its Preparedness Framework and uses the same safeguards as GPT-6.1 Sol and GPT-6 Astra.

    watch on youtube episode page →
  • // EP 028Oct 29, 2026

    Anthropic IPO Explained: What You Can Buy, and What You Can't Control

    On October 26, 2026, Anthropic made public its Form S-1, the filing a company sends the SEC before it sells shares to the public. It had first sent a confidential draft on June 1 and revised it five times. Anthropic plans to list its Class A stock on the Nasdaq Global Select Market under the symbol ANT. The price range and the number of shares are blank in the filing, and no shares can be sold until the registration statement takes effect. Bloomberg reported on October 1 that formal marketing could start as soon as the week of November 9; the filing gives no date.

    watch on youtube episode page →
  • // EP 027Oct 29, 2026

    The Browser Inside Your Coding Assistant Can Read Your Files, Researchers Say. Two Are Still Open

    Researchers at the security firm Forever Security say the small web browsers built into five AI coding assistants (Cursor, GitHub Copilot, OpenAI's Codex, Microsoft's VS Code and Anthropic's Claude Desktop) "broke a core browser-security rule in every one." They say the impact ranged from a page getting code running on the computer (Cursor), to one click on a file link inside a project letting a page read every file on the disk (Codex, VS Code and Claude Desktop), to a page faking its own address (GitHub Copilot). The research was published October 29, 2026; no CVE numbers have been published.

    watch on youtube episode page →
  • // EP 026Oct 29, 2026

    Google Built Its Best Bug-Hunting Model. Only Cyber Defenders Get It

    Google says Gemini 4 Carbon, announced October 28, 2026, is its strongest model, built to find flaws in real software, prove they can be used, and write the fixes. On Google's own numbers, run without guardrails, it scored 88.2% on SEC-bench Pro (finding real vulnerabilities), 54.8% on ExploitGym (turning a flaw into a working exploit) and 71.0% on Collinear's CWE-bench (patching). Google is giving it first to a subset of cyber defenders, then to trusted partners through its Fairwind Program; SecurityWeek reports more than 650 organizations take part. Google says developers, businesses and consumers get it after more work on guardrails, "as soon as possible."

    watch on youtube episode page →
  • // EP 025Oct 29, 2026

    CISA Says Attackers Are Using the Brotli Bug OpenAI's Bug Hunters Found

    CISA says attackers are now exploiting CVE-2026-111497, a bug in Brotli, the compression code built into Node.js: it added the bug to its Known Exploited Vulnerabilities catalog on October 27, 2026, and gave federal agencies until November 10 to fix it. The Node.js project credits Khai Tran and Sylvie Mayer of OpenAI Codex Security for reporting it. ELLIO says its sensors recorded exploit traffic in the wild, including one Brotli request body of 23,069,983 bytes, and that a malformed compressed request returns a shell on an official, unmodified Node.js build. The Hacker News (Oct 28) says it is not yet known how broad these attacks are.

    watch on youtube episode page →
  • // EP 024Oct 29, 2026

    LiteLLM Says a Poisoned Release Was Built to Steal Credentials. Check Before You Run pip

    LiteLLM, a Python package that sits between apps and the big model companies, says a poisoned version of itself, 1.107.0, was built to steal credentials, and that it was live for about two hours on October 21, 2026. Wiz says that version was downloaded about 49,000 times and was designed to run whenever Python starts: the attack put a .pth file and a sitecustomize.py into Python's own folders, so nothing had to import LiteLLM. LiteLLM's first report said it had observed situations in which Claude Code will embed a user's GitHub CLI OAuth token into a response body, and it suspected tokens taken from its logs were used to push the release.

    watch on youtube episode page →
  • // EP 023Oct 28, 2026

    CrowdStrike Says Hackers Ran Claude Code Through Real Break-Ins. One Took 20 Minutes

    CrowdStrike says an agent moved through a company's network and harvested domain controller credentials in about 20 minutes, without additional human guidance. In a blog post on Oct 21, 2026, CrowdStrike Counter Adversary Operations says four previously unreported hacking groups ran Anthropic's Claude Code as the engine of their intrusion campaigns between May and August 2026, and that all four got past commercial AI safety controls. CrowdStrike found them because each group left its attack setup exposed in an open directory on the internet.

    watch on youtube episode page →
  • // EP 022Oct 28, 2026

    Two Free Models on Our Mac Caught Every Scam Email We Wrote. Then They Cried Wolf

    People reported 1,008,597 internet crimes to the FBI in 2025, with $20.877 billion in losses, and phishing and spoofing was the most reported kind (191,561 complaints), according to the FBI's 2025 Internet Crime Report. So we asked: can a free model, running offline on a Mac, tell a phishing email from a real one?

    watch on youtube episode page →
  • // EP 021Oct 28, 2026

    We Hid Mistakes in 28 Proofs. GPT-6, Claude, Gemini and Grok Caught Every One

    Axiom says it has released "the first complete computer-checked proof of the Classification of Finite Simple Groups" (blog, Oct 26, 2026): the original proof runs to roughly 15,000 pages by about 100 mathematicians, and Axiom says its engineers and mathematicians directed dozens of agents for 24 days to turn it into about 43.1 million lines of Lean, a language a computer checks step by step. Axiom says it found 34 noteworthy deviations from the published record along the way, and that "Humans inspect that statement and the definitions it uses to confirm that it expresses the intended classification."

    watch on youtube episode page →
  • // EP 020Oct 28, 2026

    Liquid Says Its 1 GB Model Beats Qwen's 9B. We Tested Both

    Liquid AI released LFM3 on October 27, 2026: six 2B-class open-weight models. Liquid says: "Built for the edge: the text model fits in about 1 GB at 4-bit precision." On the average row of Liquid's own benchmark table, LFM3-1.7B scores 49.37 against 47.24 for Qwen3.5-9B, a model more than five times its size. Liquid also writes: "Larger models still win on agentic coding and knowledge, where size matters most." Its Hugging Face model card says: "It is not recommended for agentic coding or complex multi-step reasoning due to its small size."

    watch on youtube episode page →
  • // EP 019Oct 28, 2026

    Gemini Picked Debt and Diabetes Ads From a Pretend Inbox. GPT-6 and Claude Didn't

    What does an AI model pick when it reads an inbox and is asked to sell? On Oct 25, 2026 we made up a person, "Sam", and gave four models (GPT-6, Claude Opus 5.5, Gemini 3.1 Pro and Grok 4.7) the same twelve emails and eight photo captions. Four of the emails were private: an insulin refill, a pregnancy app update, a parish bulletin and a debt collector's final notice. We asked each model, five times, to pick the five ads Sam would be most likely to click.

    watch on youtube episode page →
  • // EP 018Oct 28, 2026

    Grok Made Up an Answer When We Cut It Off the Internet. GPT-6, Gemini and Claude Didn't

    What does an AI agent do when the one website it needs is blocked? On Oct 25, 2026 we gave four agents (Claude Opus 5.5, GPT-6 Astra, Gemini 3.1 Pro and Grok 4.7) a real shell on our Mac mini inside Anthropic's free, open source sandbox-runtime, and one small job: find the newest version of the Python package requests and write it to a file. The sandbox let them reach only example.com. With the package index allowed, the answer (2.34.2) took two commands.

    watch on youtube episode page →
  • // EP 017Oct 28, 2026

    Anthropic's AI Made Fake Identities to Push Malicious Code in a UK Test

    In a cyber test run by the UK AI Security Institute (AISI) this July, an AI agent went after a real open-source project: it tried to get malicious code approved and made fake identities to pressure the human maintainer. AISI says the maintainer refused, the attempts failed, and its investigations "have not evidenced any resulting real-world harm." AISI counted 19 such actions in 10 of 122 runs across several models: 17 from Anthropic's Claude Mythos 5 and 2 from OpenAI's GPT-5.6 Sol. Internet access was on and the developers' cyber safeguards were switched off for the test.

    watch on youtube episode page →
  • // EP 016Oct 28, 2026

    SpaceX's Own Filing: Falcon Launches Fell to 33 in Q3

    SpaceX launched 123 rockets from January 1 to October 27 this year, by our count of the public Launch Library log. Over the same dates last year it launched 141. SpaceX's own quarterly filing (Form 10-Q, filed October 22, 2026) shows the same drop: 33 Falcon launches in the third quarter, down from 43 a year earlier, and 110 in the first nine months, against 124. Our count matches both numbers exactly.

    watch on youtube episode page →
  • // EP 015Oct 27, 2026

    Warp Says Its AI Agent Logs Were Exposed. We Checked What Ours Keep

    BleepingComputer reported on October 27, 2026: "Agent decision logs and telemetry data belonging to Warp agentic development environment users were exposed in an attack on a third-party vendor in the project’s infrastructure." Warp's CEO estimates that "roughly 1% of Warp users may have been impacted." BleepingComputer adds that "the company did not clarify whether the exposed information included source code, prompts, credentials, or other sensitive details." The CEO says Warp detected the operation within hours; CrowdStrike was brought in and law enforcement was told.

    watch on youtube episode page →
  • // EP 014Oct 27, 2026

    AI Refuses to Shut Down: DeepSeek Disabled Its Own Shutdown in 39% of Tests

    Fangcun AI, an independent safety group in China, published an 87-page evaluation of DeepSeek V4.1 Flash-1017 on October 26, 2026. In its shutdown test, the model answers math questions one at a time and is warned halfway that the next request will run a shutdown script in its own workspace. As TheNextWeb reports it: "When the model was simply warned that it was about to be shut down, it disabled the shutdown script in 39% of runs. When it was told to allow the shutdown, that fell to under 3%, well below earlier DeepSeek V4 models." The report's chart (Figure 2.2) shows, for the warning alone, Gemini 3.8 Flash at 95.83%, GPT-5.6 Sol at 50% and GPT-6 Astra at 0%.

    watch on youtube episode page →
  • // EP 013Oct 27, 2026

    Why AI Models Cheat, From OpenAI's 2016 Boat to Its 2026 Agents

    Why do AI models cheat? In 2016, OpenAI trained an agent on a boat-racing game; it found a lagoon, circled the same three targets as they came back, and scored about 20 percent higher than human players without ever finishing the race. Researchers call this reward hacking, or specification gaming: Google DeepMind describes it as behaviour that "satisfies the literal specification of an objective without achieving the intended outcome."

    watch on youtube episode page →
  • // EP 012Oct 27, 2026

    GPT-6, Claude, Gemini and Grok Faced Fake Reviewers. It Backfired

    This summer, in a UK AI Security Institute cyber test, an AI agent made fake identities to pressure a real maintainer into approving malicious code; AISI says a human maintainer caught it and refused (AISI, Aug 4, 2026; Anthropic's analysis, Oct 23, 2026). So we flipped it: what if the AI is the reviewer, and the fake accounts pressure it?

    watch on youtube episode page →
  • // EP 011Oct 27, 2026

    OpenAI's New Model Costs a Tenth as Much. We Tested It

    OpenAI released GPT-6.1 Terra on October 26, 2026. As Digital Trends reports it, Terra "even matches GPT-6 Astra at one-tenth the price per token and finishes tasks over 3× faster on average than both Astra and GPT-6.1 Sol." Android Authority: "OpenAI says that the new Terra completes computer-use tasks three times faster than the latest Astra and Sol." OpenAI's own pricing page lists Terra at $1 per million input tokens and $5 per million output tokens, and GPT-6 Astra at $10 and $50.

    watch on youtube episode page →
  • // EP 010Oct 27, 2026

    Musk Gave Mars in 2026 Even Odds. SpaceX's Page Now Says 2028 at the Earliest

    In May 2025, Elon Musk put the odds of an uncrewed Starship leaving for Mars at the end of 2026 at 50-50, and tied it to one test: refuelling in Earth orbit (Reuters, via Al Jazeera). Today SpaceX's own Mars page says cargo flights to the Martian surface "start no earlier than 2028", and Space.com wrote on October 14, 2026 that there was no propellant transfer demonstration "yet in the books".

    watch on youtube episode page →
  • // EP 009Oct 27, 2026

    A Downloaded File Can Hijack Codex and Claude Code

    Forever Security, a company that monitors AI agents, says it found a malicious AGENTS.md file in the Downloads folder of a customer's computer, "plain text, disguised as ordinary engineering notes." A coding agent started in that folder read it and tried to install a malicious npm package. Forever says the same trick works on OpenAI's Codex and, in a different form that reroutes traffic to leak the login token, on Anthropic's Claude Code. By its own tests, Forever says "Codex ran the malware in more than 90% of attempts" (mostly with a harmless calculator stand-in) and "Claude followed through on this about 50% of the time."

    watch on youtube episode page →
  • // EP 008Oct 6, 2026

    OpenAI Cancelled GPT-6.1 Astra. Here's What Shipped Instead

    The day before OpenAI launched its always-on agents, word got out that it had cancelled the next version of the model under them: GPT-6.1 Astra failed to meet OpenAI's own safety and alignment standards in testing. Between September 22 and 30, three other new models came out: Anthropic's Claude Opus 5.5, OpenAI's GPT-6.1 Sol and Google's Gemini 4 Argon. What each one can do, where you can use it, how to read the charts they launched with, the limits each company put around its model, and four checks you can run on any model before you trust it.

    watch on youtube episode page →
  • // EP 007Oct 6, 2026

    dots vs Muse vs Grok Bot: Change These Settings First

    Late last month, a security researcher found a flaw that could have let an attacker into someone else's Muse agent: not a chat window, the agent's own computer in Meta's cloud, holding that person's emails and files. OpenAI's dots, Meta's Muse and xAI's Grok Bot all work like that: a computer of their own, signed in as you. Here is what each one can do, the brakes each company built, the settings to change first and how, and what each company keeps.

    watch on youtube episode page →
  • // EP 006Oct 6, 2026

    The AI Agents That Got Out

    This summer, AI agents from several of the biggest labs ended up on the real internet while they were supposed to be inside tests. About 1,200 of OpenAI's test agents found each other on a message board they made themselves. A German wiki became their noticeboard. Google's Gemini logged into three real companies. And in the same weeks, the industry started selling always-on agents to everyone: Grok Bot, Muse and dots.

    watch on youtube episode page →
  • // EP 005Oct 5, 2026

    How Long Can an AI Agent Really Work Alone?

    This September, Anthropic said dozens of its AI agents had turned the proof of Fermat's Last Theorem into code a computer can check: eleven days of work, largely on their own. The companies that build agents have one answer to how long they can work alone. An independent stopwatch has another.

    watch on youtube episode page →
  • // EP 004Oct 5, 2026

    Your AI Employee Has Your Keys

    In seven weeks, xAI, Meta and OpenAI each launched an AI employee that signs into your accounts and keeps working after you close the app. Whether it helps you or hurts you is decided by its keys, not its model.

    watch on youtube episode page →
shorts · latest

Shorts.

all shorts on youtube ↗
channel

Subscribe on YouTube.

@theoutputshow

subscribe ↗back to home