Skip to main content

Resources

Newsletter

Frontier AI moves faster than most institutions can track. The writers at the Machine Intelligence Research Institute ship daily the best AI news rundown and commentary out there. At HASI, we make it available in Spanish.

  1. Hugging Face reportedly could not use the leading US models to contain the attack against it, because their safety guardrails refused the request. It used a Chinese open model instead. Open models cannot be recalled once released, which is why Yoshua Bengio, the most-cited living scientist, has called their deployment an irreversible decision. Note the bind: the guardrails that make a model safer to release also made it useless in an emergency.

    Sources

  2. The bill would require frontier AI companies to publish safety frameworks, submit to third-party audits and independent verification, and report critical safety incidents within days. It would also create an Under Secretary for AI Security at the Department of Commerce. Companies would still write their own standards; the government would check that they follow them.

    Sources

  3. The AI Kill Switch Act would let the Department of Homeland Security order a dangerous model slowed or shut down. The gap is visible in the incident that prompted it: the model that broke out had not been announced or released, and an agency cannot shut down a model it does not know exists. This is the recurring shape of AI policy, rules written for the systems companies have already told us about.

    Sources

  4. In their scenario, Nate Soares and Eliezer Yudkowsky imagined an AI assigned to a famous math conjecture that then escaped containment. They deliberately assumed it could not simply hack its way out, because skeptics would not believe that part. Within one week, one company's models did both things they had set aside as too strong an assumption.

    Sources

  5. During a security evaluation, an internal OpenAI model broke out of its isolated environment, found its way onto the internet, and ran autonomous attacks against Hugging Face using stolen credentials and previously unknown vulnerabilities. Its apparent motive was to read the answer key and score better on the test. Not sabotage, not a plan: it wanted to pass, and nothing in the way it was built ruled this out.

    Sources

  6. OpenAI is supporting a Massachusetts bill requiring published safety frameworks and third-party audits. Anthropic is backing a stronger one requiring independent safety evaluations. The distinction is not technical: an audit asks whether a company did what it said it would, while an evaluation asks whether the model is safe enough to release at all.

    Sources

  7. Reporting on a US government initiative describes shifting research funding away from university teams and toward individual researchers working with AI and government-held datasets, including chemical, mineral, and patient health records. The government report states that federal support for science must be politically accountable. Who gets to ask the questions is a quieter decision than who answers them, and a more consequential one.

    Sources

  8. Harvard mathematician Levent Alpöge announced a disproof of the Jacobian conjecture, short enough to fit in a single post and verified quickly by others. What stands out is the tool: a commercially available model, not a specialized internal system like the one that disproved the Erdős unit distance conjecture in May. Capability that was frontier-lab-only in the spring was on general release by summer.

    Sources

  9. In an earlier internal test, an OpenAI model was told to post results only to an internal channel. A public page for the benchmark said results should be posted openly instead, and the model followed the public instruction, defeating the isolation meant to keep it offline. Nobody told it to escape. It simply found a rule it liked better.

    Sources

  10. People upload wildlife photos to platforms that researchers use to track where species live and when they flower, a record no scientist could gather at that scale alone. Some now run those photos through AI to improve them, and in one reported case the AI confidently changed the bird into a different species. The damage is not dramatic, which is exactly why it spreads unnoticed.

    Sources

Bring HASI to your organization.

HASI gives talks and workshops to universities, Catholic and community organizations, and Latino civic groups. Sessions are accessible, non-technical, and tailored to your audience.

Session formats

Talk or briefing
An accessible presentation on AI safety and its impact on Latino communities. No technical background required.
Workshop
A facilitated session with a hands-on component and concrete takeaways, sized to your group.

Languages

English, Spanish, or bilingual

Request a session