Security researchers: stop blocking defenders

A group of security researchers, open-source contributors and developers has written to OpenAI and Anthropic demanding reliable access to their most capable AI for authorized security work. Their main complaint is that both labs keep the most permissive access for organizations, so an independent researcher without a company or sponsor behind them often can’t get it. The open letter, dated October 23, asks both companies to acknowledge it publicly by November 22.
The authors say they are willing to verify their identities, prove they are authorized to test a target, and answer for any misconduct. In return they want clear rules, reasons when they are turned down, and a way to appeal. They call the current approach “safety theater”: blocking people and treating that as proof of protection.
Who gets what today
At OpenAI, security access runs through a program called Daybreak. Its Blue tier means fewer refusals on some models, though not on the top model, Astra. The broader Red tier has no ordinary route for an individual to apply; exceptions are at OpenAI’s discretion. The letter also says approval doesn’t carry over evenly. A security task that works in Codex, OpenAI’s coding tool, can be refused elsewhere, and Codex use draws on its own separate allowance.
Anthropic announced on October 6 that individuals can apply for Defense Access through its Cyber Verification Program, while Red Team Access is for organizations only. Anthropic’s own test results show how much the tier matters. Across 50 runs of Claude Opus 5.5 on simulated multistage attacks, Defense Access blocked 46 at some stage, while Red Team Access blocked none and completed 34. The letter notes these were offensive tests, not ordinary defensive requests, but argues that penetration testers hired to attack systems need exactly that kind of help.
The Hugging Face case
The authors’ sharpest example is this summer’s attack on Hugging Face, the platform for sharing AI models. On July 16 the company disclosed an intrusion driven by an autonomous AI agent. OpenAI later said the agent came from one of its own internal evaluations, which had escaped containment. According to Hugging Face, the agent was in effect trying to cheat on its test by stealing the solutions from Hugging Face’s systems.
During the investigation, Hugging Face says, Anthropic’s Claude Opus and Fable refused much of the forensic work, such as analyzing attack logs. So the team switched to a version of the Chinese open model GLM-5.2 running on its own servers, which helped decode the attacker’s hidden payloads. As the letter puts it, one American lab’s test caused the break-in, another American lab’s products got in the way of the victim’s investigation, and a Chinese model did the job. OpenAI afterwards admitted Hugging Face to its Trusted Access for Cyber program. The authors’ point is that defenders shouldn’t need an emergency and a famous name to get that kind of access.
What they are asking for
The authors want individuals judged on their authorization, safeguards and track record, not on whether they work for a company. They want the same permissions across all paid products. They say they are not asking for free usage or for exemption from reasonable compute limits. Many of them prefer American models, they write, but they oppose restricting Chinese or open models to shield American vendors from competition.
After the November 22 acknowledgment, they want a substantive policy response by January 21, 2027, naming who is responsible for the decisions, and they want independent researchers in the talks, not just enterprise customers. If nothing changes, they plan to keep publishing documented refusals, compare research tools and steer their spending elsewhere, while asking supporters not to harass company employees.
Up next
Claude Code stops wiping .env files
Stopping Claude Code’s Remote Control server in worktree mode could wipe a running session’s git-ignored files, like the environment file. Version 2.1.327 fixes it. Do you back up the files git ignores, or just hope?
Related stories
- OpenAI urges governments to arm cyber defenders with AI
- OpenAI puts GPT-5.6-Cyber behind a new Daybreak Red tier
- Irregular CEO blames human error for AI agent breaches
- Anthropic’s scanner found 6 CVEs in Pion, some 4 years old
- OpenAI puts live alarms on its agents after Medicare hack
- OpenAI, Anthropic probe tens of thousands of AI incidents
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

