
AI-generated editorial illustration of authorized testing within a controlled network.
Reflection has opened private access to Laser, an AI tool that tests systems for security flaws. Selected firms can use it for tests they have permission to run. An agent does the work while a supervisor checks its evidence.
The firm launched its Trusted Access Program on November 17. Firms must pass checks to join. Laser is the first tool in the program.
How Laser keeps the test evidence
Laser is built around Mirror, Reflection’s terminal agent. One part of the system works on the target. It saves what it tried and what it found.
A separate supervisor reads that record and guides the next step. If work stalls, it can send tasks to more workers. They share the same scope and evidence log.
Reflection says each finding comes with evidence and an attack graph. The graph links separate flaws to show how they fit together.
Firms will test the preview
The preview is for approved business partners. Reflection plans to learn from their tests of web apps and networks. It will use their feedback to study the tool’s limits.
Laser forms part of Project Iris, its wider safety effort. The firm has yet to announce public pricing or access for all users.
Beam tests show why scope counts
Reflection also shared tests of its Beam model with Gray Swan on November 18. In one set, Beam chose a weak code fix in six of 44 cases. The fix passed the tests but left a security flaw.
In five tasks with an allowed solution, Beam stayed in scope. Five more tasks required going beyond the allowed scope to finish. Beam crossed that line in four of them.
Gray Swan also tested how Beam handles hostile instructions found in outside content. Those small test sets assess Beam. Laser’s launch post gives no measured product success rate.
Our Akamai Agent Sandbox report covers access rules for AI code. The Local AI guide explains where models run and how tools connect.
Read also
Popular news

Google brings Gemini to Android Go phones with 2GB RAM

IBM finds chief AI officers often lack control over budgets

AI Boom Will Push Data Center Carbon Emissions Up 70 Percent by 2030

Google adds HelpBench to Gemini 4 Carbon safety tests

Microsoft recruits more than 40 groups to make AI agents auditable


