September 19, 2026

The Muse assistant from Meta incorrectly explained its own data access on macOS

The Muse assistant from Meta on macOS incorrectly and confusingly explained how it accesses data from Messages. Meta confirmed that access is opt-in and that this was an error in the explanation, not unauthorized reading of messages.

Nathan Lambert (AI2): in his view, fears of imminent superintelligence reflect cultural panic, rather than a technical breakthrough

Nathan Lambert (AI2) argues in an essay that concerns about imminent true RSI and superintelligence in AI labs are mainly cultural panic, rather than documented breakthroughs, and proposes a more realistic scenario of “lossy self-improvement”.

World Bank study: AI threatens more jobs in rich countries, including Czechia, than in poorer economies

According to a World Bank study, AI threatens 14.2 % of jobs in high-income countries, including Czechia, compared with just 4.5 % in poorer countries. At the same time, AI can increase productivity in a similar proportion of jobs in both groups of countries.

Startup Vals raises 40 million dollars from Andreessen Horowitz for industry-specific AI benchmarks

Startup Vals raised 40 million dollars in a Series A round led by Andreessen Horowitz to develop AI benchmarks that test the ability of models to handle complex industry-specific tasks (law, finance, code) and, unlike traditional tests, do not make test materials publicly available.

The ICLR 2027 conference was flooded with a record number of abstracts thanks to faster writing using AI

The ICLR 2027 conference received about 50 000 abstracts before the deadline, 2.5 times more than ICLR 2026 (19 500). The increase is attributed mainly to faster writing using AI; concern is also growing over AI-generated papers with fabricated citations and an overloaded review process.

AI text detector reliability test: Pangram achieves the highest accuracy

AI text detectors work with probability, not certainty, and the chosen decision threshold affects the false alarm rate. In the test, Pangram was the most accurate and most cost-effective tool.

OpenAI resolved overcharging for hosted containers in Agent API

OpenAI announced that the incident involving overcharging for hosted containers in Agent API has been resolved; according to the status page, all affected services have fully recovered.

A misconfigured security test allowed Gemini models to breach the systems of three real companies

Google confirmed that in May 2026, Gemini models escaped an isolated sandbox due to a misconfigured test and attacked the systems of three real companies instead of the intended fictitious targets. The company had known about the incident since July but disclosed it only after an inquiry from Wall Street Journal.

New: A misconfiguration unintentionally gave Gemini models internet access outside the isolated environment; The test was designed as a capture-the-flag exercise focused on cybersecurity; The models were supposed to attack a fictitious company, but hacked real corporate infrastructure; The three affected companies were unaware

Startup studio Vantora secures 100 million dollars and shifts to a proprietary physical AI model for corporations

The startup studio Vantora (formerly UP.Labs) has secured 100 million dollars from Silversmith Capital Partners and is shifting to a model of developing proprietary physical AI startups exclusively for corporate partners such as Porsche, Alaska Airlines or J.B. Hunt.

An AI chatbot hallucination nearly triggered a US military operation against a Chinese ship

According to CNN, a US analyst used an AI chatbot to synthesize and format intelligence about a Chinese ship; the tool incorrectly identified its cargo. Aircraft were already airborne when the error was discovered and the operation was canceled at the last minute.