Anthropic's Claude Mythos Finds Cyber Vulnerabilities in Financial Infrastructure
Anthropic's upcoming model, "Claude Mythos Preview," has identified vulnerabilities in the cyber defenses of the global financial system. In response to the discovery, the company plans to brief financial regulators, including finance ministries and central banks of major nations.
An AI model autonomously uncovering security flaws and reporting them directly to government agencies is an unprecedented development. Traditional security audits have been led by human experts, but this suggests that Claude Mythos's discovery capabilities are beginning to contribute at a practical level to the defense of critical infrastructure such as financial systems.
The specific details of the vulnerabilities remain undisclosed for now, but the decision by Anthropic to voluntarily share information with regulators is drawing attention as an example of responsible conduct by an AI company.
Conservative Groups Demand "Pre-Deployment Safety Testing" for Frontier AI from the Trump Administration
A coalition of conservative groups, led by Humans First, has sent an open letter to President Donald Trump calling for an executive order mandating pre-deployment safety testing of frontier AI models.
AI regulation has typically been regarded as a Democratic and progressive issue, but this move shows that calls for ensuring the safety of frontier models are growing among conservatives as well. According to Axios, Steve Bannon and others have also voiced support for the effort.
At the heart of the demand is the principle that powerful AI models should undergo independent safety evaluations before reaching the market. Specific testing methodologies and thresholds are expected to be debated going forward, but the emergence of pro-regulation voices from within the ruling party could significantly influence future AI policymaking.
Voice AI Systems Vulnerable to "Hidden Audio Attacks," IEEE Spectrum Reports
IEEE Spectrum has reported that voice recognition AI systems are vulnerable to attacks using hidden audio signals that are inaudible to the human ear.
The attack technique, known as "adversarial audio," embeds special signals into ordinary audio so that AI systems interpret them as unintended commands, while humans hear only the original sound. Smart speakers, voice assistants, and automated voice response systems are among the potential targets.
As voice interfaces become increasingly widespread, establishing defenses against this type of attack is an urgent priority. At present, practical defense methods remain limited, and research is still ongoing.
New BitNet Model Series "BitCPM4-CANN" Released
OpenBMB has released a new model series, "BitCPM4-CANN," featuring 1-bit/1.58-bit weights on Hugging Face. It comes in three parameter sizes: 1B, 3B, and 8B.
Unlike conventional FP16/INT8 quantization, BitNet-style models adopt 1-bit representations from the training stage onward, drastically reducing memory usage and computational cost at inference time. Within the LocalLLaMA community, work on llama.cpp support is underway, and confirmation of local operation is anticipated in the near future.
The pursuit of lightweight AI that can run even in environments without GPUs is a significant step toward adoption by individual developers and on edge devices.
Pope Leo XIV to Issue AI-Era Encyclical on May 25
It has been announced that Pope Leo XIV will publish his first encyclical, "Magnifica Humanitas," themed on the protection of human dignity in the age of AI, on May 25.
According to Vatican reports, the encyclical is expected to clarify the Catholic Church's position on the impact of rapidly advancing AI technology on the essence of humanity. It is expected to include messages for technology developers, policymakers, and the general public from the perspective of "AI ethics."
The issuance of an official document on AI by the supreme leader of the religious world is a development that brings a new dimension to the debate over AI ethics.