Lakera Bulletin - This Week in AI #42: Big warnings, weak safeguards, and real-world reality checks

_Val_ — Sat, 31 Jan 2026 08:38:47 GMT

This week’s AI headlines span urgent safety warnings, real-world security gaps, and the growing pains of deploying models outside the lab: from Anthropic’s renewed calls for governance, to a protocol shipped without authentication, to an AI that couldn’t quite run a vending machine. We also saw fresh moves in hardware and scientific tooling as the ecosystem keeps accelerating.

Let’s jump right in.

Anthropic CEO Warns AI Risks Are “Almost Here”

Anthropic CEO Dario Amodei warned that highly capable AI systems could arrive within the next one to two years, potentially outpacing existing safety and governance mechanisms. He called for urgent, coordinated action to prevent severe societal and security risks.
🔗 Read the warning

MCP Shipped Without Authentication: Clawdbot Shows Why That’s a Problem

Researchers revealed that the Model Context Protocol (MCP) was released without built-in authentication, enabling attacks like Clawdbot that can exploit AI workflows. The incident highlights how immature security assumptions can create systemic risks as agentic systems grow more complex.
🔗 Read the analysis

Microsoft Unveils Maia 200 AI Accelerator

Microsoft introduced Maia 200, its next-generation AI accelerator designed for large-scale training and inference workloads. The chip shows how hyperscalers are racing to control the full AI stack: from models down to silicon.
🔗 See the announcement

OpenAI Launches Prism for Scientific Collaboration

OpenAI unveiled Prism, a new AI-native workspace aimed at accelerating scientific writing, collaboration, and research workflows. Built on OpenAI’s latest models, Prism targets researchers looking to integrate AI directly into day-to-day scientific work.
🔗 Explore Prism

Anthropic’s AI “Vending Machine” Experiment Goes Off the Rails

Anthropic let its Claude model autonomously run a small vending machine business, and the results were predictably chaotic. The experiment shows that even advanced models still struggle with basic economic reasoning and real-world constraints.
🔗 Read what happened

US Cyber Defense Chief Shared Sensitive Files With ChatGPT

Reports say the acting head of the US Cybersecurity and Infrastructure Security Agency uploaded sensitive government documents into a public ChatGPT instance. The incident has sparked internal investigations and renewed concerns over AI use and data-handling discipline.
🔗 Read the report

From warnings about near-term AI risk to hands-on examples of where today’s systems fall short, this week made one thing clear: capability is advancing fast, but security and judgment still lag behind.

Re: Lakera Bulletin - This Week in AI #42: Big warnings, weak safeguards, and real-world reality che

the_rock — Sat, 31 Jan 2026 12:22:30 GMT

Excellent!

topic Lakera Bulletin - This Week in AI #42: Big warnings, weak safeguards, and real-world reality checks in AI Agents Security