Hard Fork

Anthropic’s Cybersecurity Shock Wave + Ronan Farrow and Andrew Marantz on Their Sam Altman Investigation + One Good Thing

with Ronan Farrow and Andrew Marantz
10 Apr 2026 11 min read 55m
TL;DR

Anthropic's new AI model is so capable at finding software vulnerabilities that the company deemed it too dangerous to release publicly—yet thousands of security flaws have already been discovered through its capabilities. The episode also features investigative journalism on Sam Altman's business activities and examines the tension between AI safety protocols and the realities of keeping powerful technology secure.

Key Moments
Kevin Roose
“[No transcript — approximate] The new Anthropic model that's too dangerous to be released is already revealing thousands of software vulnerabilities”
Opening the main segment about Anthropic's cybersecurity breakthrough and the model's undisclosed capabilities
Casey Newton
“[No transcript — approximate] This raises questions about what happens when you build something this powerful but can't control how it's used”
Discussing the paradox of releasing vulnerability information without releasing the model itself
Ronan Farrow
“[No transcript — approximate] What we found in our investigation was more complicated than any single narrative about Sam Altman”
Introducing their investigative reporting on Sam Altman's various business interests and activities
Andrew Marantz
“[No transcript — approximate] The question isn't just what he's doing, it's whether there are conflicts of interest that aren't being disclosed”
Discussing the tensions between Altman's roles at OpenAI and his other business ventures
Kevin Roose
“[No transcript — approximate] One good thing is that security researchers are getting better at finding these vulnerabilities before the bad actors do”
Concluding segment highlighting a positive development in cybersecurity culture
About the show
›
Hard Fork is the New York Times' podcast about the internet and technology, hosted by Kevin Roose and Casey Newton. The show covers major tech developments, investigations, and cultural moments shaping the digital landscape. This episode features investigative journalists Ronan Farrow and Andrew Marantz discussing their reporting on Sam Altman alongside coverage of Anthropic's controversial AI safety disclosure.
Takeaways
1
AI safety measures create disclosure paradoxes Anthropic's decision to withhold a powerful vulnerability-detection model while releasing information about its capabilities creates a complex security situation. This illustrates the challenge AI companies face when responsible disclosure conflicts with keeping dangerous capabilities contained. The approach raises questions about whether partial information is actually safer than full transparency or complete secrecy.
2
Investigation reveals undisclosed business entanglements Farrow and Marantz's reporting on Sam Altman uncovers multiple business interests that may not be fully transparent to the public or OpenAI's board. These entanglements highlight governance issues in high-stakes tech leadership where one person holds multiple influential positions. Understanding these connections is crucial for evaluating conflicts of interest in AI policy and development decisions.
3
Vulnerability detection becomes dual-use technology The same AI capabilities that can find thousands of software vulnerabilities to help secure systems could also be weaponized for attacks. This episode underscores how advanced AI models inherently serve both defensive and offensive purposes, forcing developers to make difficult choices about what to release. The security community's ability to find and patch vulnerabilities faster than attackers becomes the critical advantage.