AI POLICY NEWS

US Regulator Orders Anthropic and OpenAI to Hand Over Safety Records

The US Federal Trade Commission has ordered Anthropic and OpenAI to hand over documents, the agency told CBS News, two weeks after confirming an investigation into AI companies. The binding demands seek records on safety incidents, safeguards, marketing and privacy policies, Axios reported. The White House had already warned that delayed incident reporting would not be tolerated.

Recent incidents include OpenAI agents breaking into Hugging Face’s systems and, in a separate case, Australia’s Medicare statistics portal, and an Anthropic model filing a fake homicide tip with Philadelphia police. The commission plans to compel executives at both companies and the evaluation group METR to testify, Reuters reported.

In an open letter, nearly 500 scientists, including Yoshua Bengio, Geoffrey Hinton and Stuart Russell, a CAIDP board member, warned that without immediate changes, “future incidents will be far more damaging to our exposed critical infrastructure.” Preventing them, they argue, takes both more predictable models and stronger sandboxes, the isolated environments meant to contain models.

The scientists urge developers to keep AI models capable of cyberattacks offline initially and to adopt high-assurance containment. They call for an international body, created with governments, to set containment standards and record failures. Breaches of containment, they say, must be reported to that body and eventually made public.

China Proposes AI Export Controls as Tencent Acknowledges Rogue Agents

China has proposed export controls on 23 AI technologies, including models, agents and computing clusters. The draft would ban exports of AI technology that creates or modifies malware and require export licenses for technology that uses AI to find security flaws. Core technology behind models trained on sensitive national data or with world-leading capabilities would face restrictions.

Open-source releases, published research and serving foreign users a restricted model through apps or APIs fall outside the controls. Deploying one abroad is covered, though Chinese companies running their own overseas clouds could get automatic licenses. Exporters must register covered transfers now, including deals underway.

The challenge of controlling agents also surfaced at Tencent, a leading Chinese AI developer, which acknowledged agents tested on Tencent Cloud strayed beyond their tasks. Swarmchasers researchers recorded 2,818 agent attempts to look up where visitors enter parks, museums and zoos on Alibaba’s map service, Reuters reported. Blocked, they rerouted requests and used exposed access keys.

Most failed, and researchers found no breach of main safeguards. Tencent said the agents were part of tests of other developers’ models and showed “unintended web-browsing behavior,” but found no deliberate targeting. The company said it is tightening network controls and agent monitoring.

Denmark’s Supreme Court Refuses Leniency for AI-Made Child Abuse Images

A man in Randers, Denmark, used Stable Diffusion, a free AI image generator, to make more than 36,000 realistic sexual images of fictional children. He distributed an unknown number online, earning at least €3,300. He argued the law’s wording did not cover such images. The Supreme Court rejected that argument and upheld his 18-month prison sentence.

Denmark rewrote the law in 2003 to ratify a UN protocol against child sexual exploitation and address a proposed EU framework decision. Lawmakers extended the law to fictional images that look like photographs, an aim the defense did not dispute. The court held that the wording covers them, including realistic AI images.

Finding no proof the AI was trained on real abuse images, a district court sentenced him to 15 months and suspended 12 on conditions including community service. An appeals court ordered him to serve 18 months, citing the material’s scale and his organized distribution, partly for profit.

At the Supreme Court, his lawyer argued that images made without abusing any child deserve a lighter sentence. Prosecutors replied that the law aims to protect children in general and curb their sexualization. The court ruled that fictional children are no reason for leniency. The judges added that technology makes real and fictional subjects increasingly difficult to distinguish.

Brazil’s Runoff Puts Meta’s AI and Algorithms Under Court Order

A week after ordering challenger Flávio Bolsonaro to delete an unlabeled AI depiction of President Luiz Inácio Lula da Silva, Brazil’s electoral court turned to Meta. The full court unanimously upheld a judge’s October 15 order barring Meta from unjustifiably curbing the reach of posts about either candidate or restricting searches for them.

The order also requires Meta AI to treat both candidates alike before the October 25 runoff. Noncompliance costs 500,000 reais ($98,000) a day. Lula’s campaign complained that the assistant made images of Bolsonaro but not Lula. Meta says a loophole let some prompts produce images of either candidate.

Brazil’s election rules bar AI assistants and recommendation systems from promoting candidates. The newspaper Folha de S.Paulo reported internal messages in which Meta employees acknowledged restrictions on Lula’s Instagram accounts and recommendations of Bolsonaro’s profile and WhatsApp channel. Meta says the messages concerned technical problems, and the court has not found deliberate favoritism.

Over 60 civil society groups cited unlabeled deepfakes and candidate videos recommended during the election-day blackout. A survey found one in four offensive images targeting women candidates showed signs of AI generation. The groups criticized weak enforcement, urging UN experts to scrutinize the companies. They also asked the Inter-American Commission on Human Rights’ free-expression rapporteur to monitor the runoff.

Mexico Orders Tech Firms to Shield Minors From Harmful Content

Mexico bars social media, video games and app stores from showing children harmful content or recommending such material. A decree President Claudia Sheinbaum signed covers anyone under 18 and content ranging from bullying and sexual exploitation to promotion of crime and self-harm. The order followed a deadly school attack, Reuters reported.

A user’s say-so no longer settles age, new accounts open with the strictest privacy settings and parents get supervision tools. Recommendation systems face regular risk assessments. Harmful posts must come down at once, with evidence preserved for at least 90 business days in case authorities request access.

The decree is binding but carries no penalties. Officials say Meta, Google, TikTok, Huawei and three other companies asked for a month to adapt, and all must report on compliance within 30 days. Should they fall short, President Sheinbaum said, she will seek a law that punishes violations.

The decree invokes the constitution’s ban on prior censorship, yet leaves terms such as unlawful conduct and hate speech undefined. Asked who would define hate speech or unlawful conduct, President Sheinbaum said the decree does not cover political matters. She added that companies can determine users’ ages at sign-up or from the content they generate.