OASM, open-source ASM platform, puts scanners and AI agents behind approval gates

OASM's starting point is a statistic: Japan's National Police Agency counted 123 ransomware victim reports in the first half of 2026, the highest ever for a half-year, with VPN devices still the most common infection route. That is why the platform focuses on continuously inventorying what is visible from the internet — unpatched VPN appliances, open remote desktop, unmanaged subdomains — before attackers do.

The developers say the hard part was never running a scanner but running many different ones safely and continuously, normalizing their output into trackable data, and then letting AI agents work on that data without free rein. The architecture follows from that: pull-based workers that reserve execution slots before taking jobs, one separate bidirectional stream for push messages like cancellation, and scanners packaged as connectors resolved from a manifest.

Much of the design is about what the system refuses to do. Approval requests time out to denial rather than permission, a Redis outage triggers re-asking rather than assuming consent, and tools reachable without going through approval are kept read-only. The developers say this openness is intentional: an asset and vulnerability inventory is close to a perfect map for an attacker, so they want anyone to be able to verify what the tool does inside — which is also why the core is published as open source under oasm-platform/open-asm.