AI Designed Toxins Evade Screens. SAM Audio Isolates Sounds. Kiosk AI Loses Money. Benchmarks Overstate Research Skill.

Show notes

The AI news for December 27th, 2025

--- This episode is sponsored by ---

Find our more about our today's sponsor Pickert at pickert.de.

---

Would you like to create your own AI-generated and 100% automated podcast on your chosen topic? --> Reach out to us, and we’ll make it happen.

Here are the details of the day's selected top stories:

AI-generated bioweapons bypass the safety systems of DNA manufacturers.
Source: https://www.heise.de/-10794581?wt_mc=rss.red.ho.themen.k%C3%BCnstliche+intelligenz.beitrag.beitrag
Why did we choose this article?
High-risk finding with immediate policy and engineering implications: shows biosafety scanners can miss AI-designed toxin sequences. Relevant for security teams, DNA manufacturers, and regulators — calls for urgent audits, adversarial testing, and updated detection methods.

Meta's new AI model SAM Audio lets users click on tones in videos.
Source: https://the-decoder.de/metas-neues-ki-modell-sam-audio-laesst-nutzer-toene-in-videos-anklicken/
Why did we choose this article?
Practical, open-source advance in multimodal AI: applies Segment Anything-style interaction to audio, enabling intuitive source isolation and editing. Useful for developers, audio engineers, and product teams — offers code/weights to prototype new workflows and research on audio segmentation.

Anthropic experiment fails: an AI kiosk seller orders a PlayStation 5, a live fish, and incurs a $1,000 loss.
Source: https://the-decoder.de/anthropic-experiment-scheitert-ki-kiosk-verkaeufer-bestellt-playstation-5-lebenden-fisch-und-macht-1-000-dollar-verlust/
Why did we choose this article?
Concrete failure-case illustrating current limits of autonomous agents in real-world commerce. Valuable cautionary example for product teams and managers considering agentic automation — highlights need for robust guardrails, human oversight, and clearer reward/constraint design.

A new benchmark shows that LLMs still fail at real scientific research.
Source: https://the-decoder.de/neuer-benchmark-zeigt-llms-scheitern-noch-an-echter-wissenschaftlicher-forschung/
Why did we choose this article?
Important calibration for expectations: a new benchmark finds LLMs still fall short at producing original, rigorous scientific work. Useful for researchers and organizations using LLMs for discovery — suggests treating models as assistants requiring verification, experimental design, and domain expertise.

Do you have any questions, comments, or suggestions for improvement? We welcome your feedback at podcast@pickert.de.