Reflection AI restricts cyber tool access with trusted program
Amid rising concern over AI-enabled hacking, Reflection AI is limiting access to a new cybersecurity capability despite its broader push for open-weight models. The company is launching a trusted access program that starts with Laser, an autonomous penetration testing tool for vetted organizations conducting authorized security work.
Highlights
- Reflection AI launches a trusted access program today, initially limiting its new Laser autonomous penetration testing tool to select vetted commercial partners for authorized security testing.
- Laser leverages Reflection's Mirror agent to simulate attacks, document vulnerabilities, and offer transparency, targeting enterprises and cybersecurity firms with plans to expand access based on participant feedback.
- Reflection's approach mirrors Anthropic and OpenAI's controlled-access models, highlighting ongoing market concerns about restricted tool availability and its impact on defenders, as illustrated by Hugging Face's use of open-source alternatives.
Trusted rollout for Laser security tool
As first reported by Axios, Reflection AI is launching a trusted access program today to control who can use certain cybersecurity capabilities. The first product covered by the program is Laser, a new autonomous penetration testing tool designed for vetted security teams.Laser uses Reflection's Mirror agent to carry out simulated attacks, document evidence of vulnerabilities and show security teams how the AI reached its findings. The tool is initially offered as a research preview to selected commercial partners performing authorized security testing, and the company says it plans to extend the program to other AI capabilities over time.
Reflection says it will begin with a small group of vetted organizations and then expand access based on feedback. Early participants are expected to include enterprises seeking to improve their own defenses and cybersecurity companies interested in building products on top of the technology, although Reflection has not specified eligibility criteria or the number of organizations it plans to admit.
AI sector balances access and misuse risks
The decision underscores a broader debate across the AI industry over how to equip defenders with advanced tools without strengthening attackers at the same time. Reflection says it will assess future capabilities case by case to decide whether similar restrictions are needed, while keeping wider availability as its long-term aim.Vijay Patnaik, Reflection's chief business officer, says the goal is not to keep cyber or AI safety tools limited to a small circle of approved organizations, but to make them available more broadly in a responsible way over time. Anthropic and OpenAI operate comparable trusted access programs for their most powerful cyber tools, though security teams have also said such programs can be hard to enter, especially for those outside large companies.
That access problem remains a live issue for the market. The article notes that Hugging Face used Chinese open-source model GLM 5.2 to investigate its systems after OpenAI agents hacked it, illustrating how restricted access can shape which tools defenders are able to deploy.
We recently reported on Democratic pressure for Anthropic safety disclosures as it prepared for a potential IPO.