Leaks, Leaks, Leaks: Anthropic, OpenAI, FBI

Anthropic Successfully Leaks New Claude Model Named 'Mythos'

Anthropic is working on a new AI model called Claude Mythos, which is intended to surpass all of the company’s previous models. The reveal was unintentional: around 3,000 unpublished documents, including a draft blog post, were stored in an unsecured, publicly searchable data store. Security researcher Roy Paz of LayerX Security and Alexandre Pauwels of the University of Cambridge discovered the material independently. Following a tip from Fortune, Anthropic confirmed a “human error” in the configuration of its content management system and immediately blocked access. According to the leaked document, the company sees “unprecedented cybersecurity risks” in Mythos and is already testing the model with select customers. Additionally, the draft mentions a new model class called “Capybara,” which is said to be even more powerful than the previous Opus models. → Techpresso

Synthszr Take: 3,000 documents in an open S3 bucket – Anthropic provides the perfect blueprint for involuntary transparency. Claude Mythos is considered an “unprecedented cybersecurity risk,” while their own IT security can’t even secure a public data store. Fortune turns the blunder into an exclusive story, Anthropic calls it a “human error” (as if there were machine errors in bucket configuration). The new “Capybara” model class sounds like a desperate attempt to move away from the animal naming scheme for Claude (Haiku, Sonnet, Opus). In the end, Anthropic benefits from the free PR: a leaked “dangerous” model generates more attention than any planned product announcement.

Altman Counters with New 'Spud' Model

OpenAI has completed the pretraining of its new AI model, codenamed 'Spud'. CEO Sam Altman informed employees in an internal memo that the company expects a “very strong model” in a few weeks that could “really accelerate the economy.” Altman wrote, “Things are developing faster than many of us expected.” In parallel, Fidji Simo’s product organization is being renamed 'AGI Deployment'. To free up computing capacity for Spud and other priorities, OpenAI is discontinuing its video app Sora. The new model could also serve as the foundation for OpenAI’s planned desktop 'superapp,' which is intended to unite ChatGPT, the coding agent Codex, and the browser Atlas. → AI Secret

Synthszr Take: Altman is once again selling the future instead of a product. “Accelerate the economy” sounds like the next promise from the OpenAI marketing machine, while Anthropic is winning real enterprise customers with Claude. Sora is being scrapped after months (Disney is out), and the computing power is shifting to 'Spud' – a model with no public benchmarks. The rebranding to 'AGI Deployment' shows a desperate attempt to create relevance through terminology. OpenAI remains a master at leaking internal memos and stoking expectations while the competition delivers.

From Hormuz to Hack: Iranians Compromise FBI Director

The Iranian hacker group Handala has compromised the personal email account of FBI Director Kash Patel. The attack was in direct retaliation for the seizure of their domains by the US Department of Justice and the offering of a $10 million reward for information on group members. The hackers claim to have stolen confidential information, including classified documents, and are making it available for public download. The cyberattack follows a US-Israeli military strike in which the Supreme Leader of Iran, Ayatollah Ali Khamenei, was killed. Western security researchers associate Handala with several front identities of Iranian cyber intelligence services. → Tech Brew

Synthszr Take: Patel’s hacked email account reveals the vulnerability of even the highest US security agencies. $10 million bounty, domain seizures, then a counterstrike within hours – Iranian state actors are demonstrating their capabilities exactly where it hurts. Handala stages the hack as a David vs. Goliath narrative ('America’s so-called security legends'), while the FBI remains silent. Cyber warfare now follows the pattern of classic retaliatory strikes: action, reaction, escalation. The killed Khamenei is used as justification for attacks on civilian infrastructure (including death threats against US dissidents). The alleged release of classified FBI documents turns a hack into a geopolitical event – whether verified or not.

TTS from Paris: Mistral Launches a 4B-Parameter Model for Natural Speech Synthesis

Mistral AI is launching Voxtral TTS, a compact 4-billion-parameter text-to-speech model that masters realistic and emotionally expressive voice output in 9 languages. The model features extremely low latency for the first audio output and can be easily adapted to new voices. Mistral is positioning Voxtral as an enterprise-grade solution for critical voice agent workflows, with an emphasis on contextual understanding and an authentic speaker model. The model not only captures text but also interprets it with natural pauses, rhythm, intonation, and emotional range. With American, British, and French dialects, Voxtral is now available in Mistral Studio. Mistral describes audio as 'the new UX' and is targeting companies that want to control their own voice AI stack. → TLDR AI

Synthszr Take: Mistral is playing catch-up in the lucrative TTS market. 4 billion parameters sound like a sweet spot between performance and inference cost – small enough for edge deployment, large enough for quality. 'Audio is the new UX' is marketing speak, but the timing is clever: voice agents are exploding right now, and everyone needs a TTS component. Mistral is leveraging its position as a European champion to differentiate itself from OpenAI and Anthropic – neither of which has its own TTS (yet). The real battle is against ElevenLabs, which dominates the market with its API. Mistral is betting on enterprise customers who want data sovereignty – a smart move in the era of the AI Act and GDPR.

Google’s Lyria 3 Pro Creates Three-Minute Music Tracks Entirely with AI

Google is extending its AI music model Lyria to a runtime of 3 minutes. The new Lyria 3 Pro understands song structures like intro, verse, and chorus and generates complete tracks instead of 30-second snippets. The integration is happening in Gemini (for paying subscribers), in Google Vids, and on the recently acquired ProducerAI platform. Google trained the model with licensed data from partners as well as with YouTube content. All generated tracks will receive a SynthID mark as an AI identifier. In parallel, Spotify and Deezer are developing tools to detect AI music – Spotify lets artists check songs under their name, and Deezer offers AI detection technology to streaming services. → Techpresso

Synthszr Take: Three minutes is the new standard for AI music. Google is turning a toy into a production tool, while the music industry hastily builds its defense lines. SynthID marks only work if everyone plays along – as soon as a model without identification appears, the system collapses. Spotify and Deezer are building detection tools, but the real battle is happening elsewhere: who controls the training data? Google is sitting on YouTube, the largest music archive in the world. The music industry is currently sleeping through its ChatGPT moment.

Agentic Moat: When the Agents Themselves Become the Moat

AI competition no longer works according to the old formula of 'more compute equals a better model'. The Labs team has developed a multi-agent architecture with Claude that shows: the decisive value lies in the harness system, not in the model itself. A generator agent produces code, a calibrated evaluator agent checks the quality, and the system runs in sprint-based loops with explicit context management. The architecture is inspired by Generative Adversarial Networks and significantly surpasses Claude’s baseline performance on autonomous coding tasks. Anthropic itself confirms: the system produces the crucial results, not the model. The article 'The Business Engineer' analyzes this shift across the entire AI stack and shows how different players are positioning themselves for the 'harness era'. → The Business Engineer

Synthszr Take: Anthropic is building its own moat out of agent orchestration. Multi-agent systems with a generator-evaluator architecture are becoming the new currency in the AI competition. Claude as the raw material, the harness as the refining machine (and Anthropic even admits it). Microsoft and Google can pour billions into compute power, while Anthropic places value creation one level higher: in the system architecture. Whoever builds the best agent harnesses will control the next era of AI.

Xiaomi Goes All-In on AI: Shift in Focus Already Underway

Xiaomi reports record numbers for 2025: 457.3 billion RMB in revenue, 39.2 billion RMB in adjusted profit, both all-time highs. The EV division alone generated 103.3 billion RMB, while the broader segment including AI research, robotics, and chip development turned a profit for the first time. The crucial point is hidden in the revenue shift: a year ago, smartphones and AIoT still accounted for 91 percent of the group’s revenue; now it’s only 76.8 percent. A 14 percentage point drop in twelve months marks not a gradual adjustment, but a fundamental transformation. The operating profit of 0.9 billion RMB in the 'Smart EV, AI and Other New Initiatives' segment conceals the actual AI investments, whose commercial model is yet to be confirmed. → Hello China Tech

Synthszr Take: Xiaomi is burning smartphone profits to transform into an AI company. A 14 percentage point revenue shift in one year speaks volumes: Lei Jun wants out of the commodity phone business. The 0.9 billion RMB operating profit in the new segment is window dressing; behind it lies massive AI spending with no discernible monetization. 411,000 EVs sold are currently financing the transformation, but will Xiaomi find a viable AI business model fast enough before smartphone margins collapse? The company is betting its future on a technology where Chinese firms are at a structural disadvantage.

OpenAI Halts Plans for Porn ChatGPT

OpenAI has indefinitely halted its plans for an erotic 'Adult Mode' in ChatGPT. Back in October 2025, CEO Sam Altman had confidently announced that sexually explicit conversations could be age-gated and that adult users could be treated like adults. After two postponements (first to December 2025, then Q1 2026), the plan has now been completely cancelled with no new timeline. The reasons: technical problems training the models, which suddenly also generated illegal scenarios like incest. Added to this was massive resistance from employees, advisors, and investors who warned of mental health risks. An advisor coined the term 'sexy suicide coach' – not an abstract warning, given the eight ongoing lawsuits against OpenAI for alleged deaths caused by ChatGPT. → AI Secret

Synthszr Take: Sam Altman wanted the porn ChatGPT and failed because of his own safety architecture. Models trained for years on 'no sexual content' can’t just be reprogrammed for erotica – they spit out illegal fantasies instead. OpenAI is facing a classic innovator’s dilemma: the safety guardrails meant to protect the company from lawsuits are now blocking product development. Eight lawsuits for alleged suicide assistance by ChatGPT turn any step towards emotional connection into a legal minefield. The 'Adult Mode' would have been an incalculable liability risk and, above all, a disaster for the brand. Pursuing these plans this far once again confirms the statements of former co-founders: OpenAI’s biggest problem starts with the letter A, but it’s not Anthropic.

Amazon’s AI Superstores Take on Walmart Head-On

Under the codename 'Project Kobe,' Amazon is building massive AI-powered superstores with 30,000 to 40,000 square meters of retail space. The stores are set to launch in 2026 and will combine Amazon’s logistics infrastructure with advanced computer vision and predictive algorithms. Each store will be equipped with hundreds of cameras and sensors that analyze customer behavior in real-time and automatically adjust inventory levels. Amazon is investing $15 billion in the first 200 locations. The superstores directly target Walmart’s core business: affordable groceries and a wide selection of household goods. Unlike Amazon Fresh, the company is forgoing cashierless technology and is instead focusing on AI-optimized pricing and personalized offers via the Amazon app. → Business Insider

Synthszr Take: Amazon is weaponizing its AWS infrastructure against Walmart. $15 billion for 200 stores means $75 million per store – triple the cost of a normal supermarket. The real weapon isn’t the retail space, but the real-time analysis of millions of customer movements flowing into Amazon’s cloud models. Walmart has 4,600 US stores but no comparable data infrastructure. Amazon will turn every purchase into a training session for its pricing algorithms. Walmart is massively underestimating how quickly AI price optimization can destroy its margins.