KI-Update – ein heise-Podcast · Isabel Grünewald, heise online

KI-Update kompakt: Anthropics Mythos, Leben mit Superintelligenz, Meta, Waymo

·22 min·1 clip

Anthropic's Mythos model found a 27-year-old vulnerability in OpenBSD and a 17-year-old bug in FFMPEG missed by millions of automated security scans, then combined individually minor flaws to enable full Linux server compromise.

1. KI-Update is a German AI news podcast from heise online, hosted by Isabel Grünewald with editorial support from The Decoder; this compact episode covers approximately ten stories from April 8th 2025. 2. The episode features three named correspondents: Martin Holland (heise online newsroom) on Anthropic Mythos, Kim Scheurenbrand (The Decoder) on OpenAI's policy paper, and Robin Ahrens (c't Redaktion) on Meta's AI models, plus Marco Pauli and Ben Schwan. 3. The first story centers on Anthropic's Mythos AI model, described as exceptionally capable at both finding and exploiting software vulnerabilities, including the ability to chain individually low-risk flaws into combined attacks that can fully compromise Linux servers. 4. Mythos found a 27-year-old vulnerability in OpenBSD — known for its security — and a 17-year-old bug in FFMPEG that had survived millions of automated security scans, in addition to thousands of other vulnerabilities across all major operating systems and browsers. 5. Anthropic's Project Glasswing gives selected IT companies and organizations including the Linux Foundation access to Mythos plus $100 million in tokens to patch discovered vulnerabilities before competing AI models develop equivalent capability, framing the effort as a 'race against time.' 6. The model will not be released publicly; Anthropic offered the US government a review opportunity, though the segment notes the White House and Pentagon have classified Anthropic as a supply chain risk, creating an ongoing conflict. 7. The EU's three main institutions — the European Commission, Parliament, and Council of Ministers — have banned their press teams from using fully AI-generated images and videos in official communications, citing the need to maintain public trust through authenticity. 8. OpenAI published a 12-page policy paper on the transition to superintelligence, presented as a discussion document rather than a formal demand; its proposals include a public wealth fund distributing AI-driven economic gains to all citizens regardless of income or wealth. 9. The OpenAI paper proposes piloting a four-day workweek at full pay in employer-union negotiations, making it permanent if productivity is maintained, and automatic social safety net measures triggered by specific labor market warning indicators. 10. OpenAI calls for formal worker co-determination rights over AI deployment at their workplaces, arguing workers best understand their own jobs and should decide where AI is introduced first — with dangerous or monotonous tasks as the priority. 11. The paper explicitly identifies OpenAI itself as a concentration risk and proposes governance audits to prevent any single person or internal group from using AI systems to accumulate power. 12. An analysis commissioned by the New York Times from AI startup UMI tested over 4,300 Google searches: Gemini 3 answered correctly in 91% of cases (up from 85% with Gemini 2), but 56% of correct answers could not be verified by the linked sources (up from 37%). 13. Facebook and Reddit appeared as the second-most-cited sources in Google's AI Overviews, which the segment notes may reflect Google's interest in citing sources less likely to sue over content usage. 14. Meta plans to release new AI models including some fully open-source; the most capable models will remain closed while smaller models will be fully open-source — a hybrid strategy attributed to new Chief AI Officer Alexander Wang, appointed mid-2024 after Llama 4 underperformed competitors. 15. Internally, Meta has a leaderboard called 'Claudonomics' tracking token consumption across 85,000 employees; the top individual consumed an average of 281 billion tokens over 30 days, and Jensen Huang of Nvidia stated he would be 'deeply concerned' if an engineer earning $500,000 did not consume at least $250,000 worth of tokens annually. 16. Intel joined Elon Musk's TeraFab project alongside SpaceX, xAI, and Tesla to build a massive semiconductor complex targeting one terawatt of annual AI accelerator computing capacity. 17. Anthropic signed a major computing deal with Google and Broadcom for 3.5 gigawatts of TPU capacity starting 2027; Anthropic's annualized revenue reached $30 billion, up from $9 billion reported at end of 2025. 18. George Hotz's TinyCorp released GPU drivers enabling external Nvidia/AMD graphics cards via Thunderbolt/USB-4 enclosures on Apple Mac computers, initially working only with his TinyGrad framework and enabling models like Qwen 3.5. 19. AI music generator Suno's copyright protection was found to be easily bypassed by small modifications — half or double playback speed, brief noise at start or end — and the service only checks content at upload, not export, allowing generated songs to be uploaded to streaming services without involving original rights holders. 20. Waymo began mapping London streets ahead of a planned public robotaxi launch later in 2025; no commercial robotaxi service currently operates anywhere in Europe, and success in London's historically complex street network would accelerate continent-wide autonomous vehicle expansion.

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from KI-Update – ein heise-Podcast