Exclusive: Reflection AI gates new cyber capabilities
Add Axios as your preferred source to
see more of our stories on Google.

Illustration: Annelise Capossela/Axios
One of the AI industry's biggest advocates for open models is taking a more cautious approach to cybersecurity and restricting access to a new security tool amid growing concerns about AI-powered hacks.
Why it matters: The move highlights a growing tension in the AI industry between making powerful models widely available and preventing their most dangerous capabilities from falling into hackers' hands.
Driving the news: Reflection AI, which has built its business around developing open-weight models, is launching a trusted access program today that will limit access to cybersecurity capabilities to vetted organizations, the company first shared with Axios.
- The program is starting by gating access to Laser, a new autonomous penetration testing tool that will allow vetted security teams to use AI agents to identify and verify vulnerabilities in their applications and networks.
- Laser uses Reflection's Mirror agent to conduct simulated attacks, recording evidence of its findings and allowing security teams to review how the AI reached its conclusions.
- The tool is initially available as a research preview to select commercial partners conducting authorized security testing, with plans to expand the program to other AI capabilities in the future.
Zoom in: Reflection plans to initially give a small group of vetted organizations access to Laser before expanding the program based on their feedback.
- The first participants will include enterprises looking to strengthen their own defenses and cybersecurity companies interested in building products using the technology.
- Organizations will need to be vetted and use Laser only for authorized security testing, though Reflection has not detailed its eligibility requirements or how many participants it plans to accept.
- Reflection will evaluate future AI capabilities on a case-by-case basis to determine whether they warrant similar restrictions, with the ultimate goal of making its tools more widely available.
The big picture: AI companies are reckoning with the best way to get advanced AI tools in the hands of defenders without also giving hackers an edge.
- Anthropic and OpenAI maintain similar trusted access programs to keep adversaries away from their most powerful cyber tools.
- "The goal is not to gatekeep cyber or AI safety to a small set of authorized organizations," Vijay Patnaik, chief business officer at Reflection, told Axios. "All of this is in service of figuring how this can be made available to defenders more broadly over time, but we want to do that responsibly."
Yes, but: Many defenders have hit hurdles while navigating these trusted access programs, noting that access is sometimes difficult to get unless someone works for a major company.
- Hugging Face had to use Chinese open-source model GLM 5.2 to investigate its systems after OpenAI agents hacked it.
Go deeper: Anthropic and OpenAI are now cybersecurity's kingmakers
