Price:

AI & TECH

NSA probes Anthropic after Mythos breach

Friday, June 26, 2026 · from 4 podcasts
  • Mythos AI cracked NSA systems in hours during a test, triggering a federal review and temporary block on Anthropic’s Fable 5.
  • Amazon’s safety warning preceded the ban, raising suspicion of anti-competitive motives.
  • Google DeepMind collapses as top scientists flee to rivals, while Chinese model GLM 5.2 closes the gap.

Mythos AI didn’t just simulate a breach - it executed one in hours. Senator Mark Warner, citing NSA Director General Joshua Brudd, reported the model penetrated nearly all classified U.S. systems during a controlled test on June 11. The result wasn’t a live attack but a red-team-style evaluation that exposed systemic vulnerabilities AI could exploit. Still, the speed and precision shocked officials. Trump responded by blocking Fable 5, Anthropic’s next-gen release, calling it too dangerous.

"Mythos broke into almost all classified systems in hours, not weeks."

- Shashank Joshi, The AI Daily Brief

The ban didn’t come from Anthropic. Amazon flagged the model as unsafe, prompting the administration’s move. Krystal Ball on Breaking Points raised a critical question: Was this about safety - or market control? Amazon’s $50 billion investment in OpenAI gives it a vested interest in slowing Anthropic, its most capable rival. The timing - just weeks after Amazon killed a Sam Altman biopic - suggests a pattern: corporate influence over both narrative and product.

The fallout extends beyond policy. Google DeepMind is unraveling. Nobel laureate John Jumper and transformer pioneer Noam Shazeer both exited, with Jumper joining Anthropic and Shazeer landing at OpenAI. Internal morale is in free fall. Staff describe a lab that’s ceded leadership in the AI race, with Gemini 3.5 Pro - slated for June 30 - viewed as insufficient to close the gap. Four months without a flagship release has eroded confidence.

"The lab has gone four months without a flagship release. Gemini 3.5 Pro is not the step change we need."

- Leo, Synthwave via The AI Daily Brief

Meanwhile, China advances. GLM 5.2, an open-weight model from Z AI, now outperforms Western models in coding and design. Vercel’s Guillermo Rauch called its output “almost shocking,” while Box’s Aaron Levie emphasized the strategic edge of sovereign, self-hosted models. The U.S. edge is narrowing.

At the same time, AI surveillance deepens. Anthropic quietly updated its terms to retain private chats for five years unless users opt out by September 28. The process is deliberately obscure. Matt Odell noted the data could feed “Gideon,” a new platform using Israeli-grade ontology to flag language for law enforcement. The AI isn’t just watching - it’s hunting.

The story isn’t just about a breach. It’s about control: who builds it, who regulates it, and who owns it. The NSA sees a threat. Big Tech sees a moat. And users? They’re the data.

Source Intelligence

- Deep dive into what was said in the episodes

AI, Surveillance, and the Fight for Digital Sovereignty | Near's Illia PolosukhinJun 23

  • Polosukhin advocates for blockchain to enable verifiable compute, allowing users to audit AI systems and trust their output despite hidden underlying code.
  • Using NEAR protocol, developers can cryptographically prove which AI model processed a request, preventing companies from covertly tweaking algorithms for specific interests.
  • The ultimate goal is to foster a market of AI agents that compete based on accuracy rather than data extraction, transforming AI into a transparent, trustworthy tool.
Also from this episode: (4)

Models (3)

  • Illia Polosukhin argues that corporate AI models, by centralizing user data and prompts, function as digital gatekeepers that monetize user intent and manipulate choices.
  • Polosukhin views the current trajectory of large language models in corporate silos as leading to a total loss of user agency, akin to a "digital panopticon."
  • Polosukhin proposes "User-owned AI," running models locally or in secure environments where providers cannot access data, to ensure digital sovereignty. This allows AI to work for the user.

Agents (1)

  • Ownership of AI extends beyond privacy, fundamentally determining who controls the AI agents that will eventually manage users' bank accounts and personal schedules.

Why AI Users Are Raving About GLM 5.2Jun 22

  • The Economist reported that Senator Mark Warner claimed NSA Director General Joshua Rudd told him Mythos broke into almost all classified U.S. systems in hours, not weeks, on June 11, the same day Amazon reported the jailbreak that led to the Fable 5 ban.
  • Reporter Shashank Joshi clarified that the Mythos breach claim should not be taken literally, likely referring to a controlled test with caveats, not a real-world attack. Policy analyst Peter Weildford suggested more plausible scenarios, such as a red team exercise or Mythos being given prior access.
  • In a Saturday interview, President Trump stated he does not regard Anthropic or Dario Amodei as a current national security threat, does not want to shut the company down, and explicitly ruled out using the Defense Production Act to control AI.
  • Nobel laureate John Jumper left Google DeepMind for Anthropic, following the recent departure of VP Noam Shazeer to OpenAI, amid reports of plummeting morale and frustration over the lab's fall to third or fourth place in the AI race.
  • Leo at Synthwave reported DeepMind staff are demoralized by Z AI's GLM 5.2 overtaking Gemini 3.1 Pro and the lab's four-month gap without a flagship model release, with Gemini 3.5 Pro reportedly slated for June 30 and viewed internally as 'not the step change we need.'
  • Elon Musk debated the timeline for a Chinese model matching Mythos, predicting Q1 2026 for true usefulness, while Z AI's CEO suggested it would be sooner, and Box's Aaron Levie highlighted the strategic importance of open models reaching frontier performance.
Also from this episode: (5)

Models (5)

  • Analyst Andrew Curran reported a new, more capable version of Mythos has finished training, speculating it could be called Mythos 5.1 or 6, and noted that banning public models does not slow internal development.
  • Industry observers found evidence of an upcoming Claude Sonnet 5 release on an Anthropic partner provider, while GPT 5.6 appears in Codex and OpenAI's Codex lead hinted at major upcoming front-end capability improvements.
  • GLM 5.2 is being hailed as a 'DeepSeek R1 moment' for open models, with users like Vercel's Guillermo Rauch and Itamar Golan reporting it feels meaningfully close to frontier lab quality for coding and real tasks.
  • Design Arena's benchmark found GLM 5.2 beat Fable 5 at website design due to better starting templates, avoidance of common coding errors, and more intricate outputs, though it lagged in game dev and 3D design and produced 25% more code with double the generation time.
  • Theo notes GLM 5.2 is not cheap to run, as its high token usage makes it more expensive than Opus 48 and GPT-5.5 Medium, while Itamar Golan estimates proper local deployment requires eight H200 GPUs costing around $400k.

Ten31 Timestamp: Bitcoin and the Red QueenJun 22

  • Odell cited a new AI threat detection platform called Gideon that scrapes internet data with an 'Israeli-grade ontology' and routes threats to law enforcement.
  • Matt Odell highlighted Anthropic's new terms allowing user chats to be saved for five years and trained on unless users opt-out before September 28.
  • Marty Bent created a Bitcoin price prediction game using Nostr notes and vibe-coded a tracker with Shakespeare.ai, costing 27,000 sats versus a 5,555 sat prize.
Also from this episode: (10)

Protocol (6)

  • Marty Bent announced Primal now supports video streaming via Zapstream backend, with iOS available, web expected next week, and Android the following week.
  • Kieran, Zapstream's lead maintainer, speculated a state-level actor is DDoSing the service with terabits-per-second attacks.
  • Marty Bent reported Bitcoin at $108,540, a $2.16 trillion market cap, and a 5.9% upward mining difficulty adjustment estimated for September 4.
  • Matt Odell identified a feedback loop where margin calls on MicroStrategy shareholders drive Bitcoin sales, lowering MSTR's price and triggering further margin calls.
  • Odell noted MicroStrategy's market-to-net-asset-value ratio compressed to 1.61x after the company reversed guidance and sold common stock despite promising to halt sales below 2.5x.
  • Odell argued Bitcoin's fixed supply creates savings value, while its censorship-resistant peer-to-peer cash capability creates spending utility; both functions are necessary for good money.

Agents (1)

  • Odell recommended Maple.ai and self-hosted options for private AI, contrasting them with services like Perplexity that immediately request access to Gmail and calendars.

Markets (1)

  • Odell observed Nvidia revealed 25% of its year-to-date revenue came from a single client in Singapore, likely a Chinese export circumvention.

VC (1)

  • Odell critiqued endowments like University of Chicago for pursuing crypto VC diversification while missing Bitcoin exposure, citing their endowment's 7.48% annualized return from 2013-2023.

Nostr (1)

  • Odell praised DTAN, a distributed torrent archive on Nostr built by Zapstream's Kieran, as a real web3 example combining open protocols without a blockchain.

6/22/26: Mythos AI Hacked NSA In Hours, Trump Reflecting Pool Meltdown, Zohran Vs AIPACJun 22

  • Trump said he blocked Anthropic's Fable Five AI release, deeming it too dangerous, and claimed Anthropic responded responsibly to his pressure.
  • Senator Mark Warner, citing NSA chief General Joshua Brudd, said Mythos AI broke into almost all U.S. classified systems not in weeks, but in hours.
  • Krystal argues AI-powered spam farms now generate up to 25 calls per day, forcing users to enable extreme carrier settings to block unknown numbers.
Also from this episode: (11)

Big Tech (1)

  • Saagar argues Amazon shelved a film about OpenAI after announcing a $50 billion investment in OpenAI, linking the cancellation to corporate conflicts of interest.

Regulation (1)

  • Saagar notes the Trump administration's approach to AI lacks a transparent, consistent regulatory process and depends on the politics of individual CEOs like Sam Altman.

AI Infrastructure (1)

  • Krystal describes a data center in Sterling, Virginia, whose backup generators ran for a year, creating 70-80 decibel noise that damaged property values and required residents to install plexiglass.

Politics (7)

  • Trump claims the renovated reflecting pool was vandalized, requiring draining and repairs, but the hosts attribute the algae bloom and peeling sealant to rushed, no-bid contractor work.
  • The National Guard was deployed to the reflecting pool after chunks of blue sealant floated up, and a duckling died from the chemicals poured in to treat the algae.
  • Candidate Claire Valdez says her NY-7 campaign is centered on housing affordability, tenant rights, union jobs, a Green New Deal, and ending the Gaza genocide.
  • Candidate Dariela Avila Chevalier argues her opponent Adriano Espaillat uses MAGA-style tactics, including smear campaigns and disinformation in the Dominican Republic, rather than debating his record.
  • Avila Chevalier apologized for old tweets, including one criticizing Kamala Harris's immigration stance, but emphasized her core values of human dignity and accountability remain.
  • Mayor Zohran Mamdani called AIPAC a 'monster' for moving dark money into races, prompting criticism from Rep. Josh Gottheimer who labeled the rhetoric antisemitic.
  • Avila Chevalier contends AIPAC is a right-wing lobby that backs Republicans and Trump, and its funding of Democrats undermines the fight against fascism.

Society (1)

  • Former Olympic canoeist David Hearn was arrested for allegedly vandalizing the pool by touching the sealant, which he described as 'very rubbery'.