Anthropic’s AI Model Mythos Sparks Safety and Transparency Concerns

Anthropic’s new AI model, Mythos, raised safety and cybersecurity concerns. Matt Shumer, the co-founder and CEO of OthersideAI, discussed these issues on ‘The Sunday Briefing.’ Mythos has the ability to find software vulnerabilities autonomously. This capability prompted Anthropic to restrict public access and provide early access to key companies and the US government for cyber defense.

Shumer highlighted that the proactive measure aims to protect critical infrastructure and national security from future threats.

HalluSquatting becomes more dangerous when an AI agent can download files and run terminal commands without close supervision.

The Tarbell Center for AI Journalism places fellow reporters with various media outlets like Bloomberg, Time, NBC News, The Guardian, and the Los Angeles Times. These fellowships are funded by donor-supported AI journalism groups, leading to questions about disclosure standards in news reporting.

Not all media outlets disclose these funding relationships, raising ethical questions. Fellows are paid between $60,000 and $80,000, with experienced fellows earning up to $110,000.

Tarbell emphasizes its commitment to independent journalism and states that it does not control the editorial decisions of the reporters it supports. The organization stresses that funding sources, like Coefficient Giving and the Future of Life Institute, do not influence journalistic content.

Certain donors, such as Facebook co-founder Dustin Moskovitz, support both Tarbell and Anthropic, complicating transparency in reporting on AI companies.

For example, some fellows wrote extensively on Anthropic without disclosing their funding source. An article in Time referred to Anthropic’s Claude model as ‘world-class’ but did not mention Tarbell’s support.

Some news outlets, like NBC News, do include editorial notes clarifying the involvement of Tarbell-supported fellows. This ensures transparency and maintains journalistic integrity.

Tarbell urges transparency among its fellows and their host publications. It requires author bios to include acknowledgments of support during active fellowship cycles. However, once a fellow is hired full-time by the host publication, disclosure may end.

Transparency and safety remain pivotal as AI models like Mythos evolve, stirring both hope and caution within the industry.