
open-asm is a self-hostable, open-source Attack Surface Management platform that inventories internet-facing assets, runs distributed scans via pluggable connectors, and tracks findings for authorized security teams.
| Tool | oasm-platform/open-asm — open-source, self-hosted Attack Surface Management (ASM/EASM) platform with AI agent integration |
| Category | Attack surface management / external asset discovery and vulnerability tracking platform |
| Primary Use | Maintaining a unified inventory of domains, IPs, ports, services, TLS certificates and URLs, then running scheduled, connector-based scans to normalize findings into trackable vulnerabilities |
| Safe Use | Intended for defenders and authorized security teams inventorying and assessing attack surfaces they own or are contractually permitted to test; all data stays on the operator's own infrastructure |
| Telemetry Note | Because scanning originates from the operator's own distributed workers, targets will see scan traffic from those worker IPs; defenders can detect connector activity (e.g. nmap, katana, ffuf signatures) against their own monitored estates, and OASM itself keeps a JSONL audit log and per-conversation AI usage records |
open-asm is a self-hostable, open-source Attack Surface Management platform written in TypeScript, currently sitting at 277 stars under the oasm-platform organization and licensed under GPL-3.0. The premise is straightforward and well-articulated in the README: external attack surfaces grow faster than teams can track — forgotten subdomains, exposed ports, outdated frameworks, leaked URLs, misconfigured cloud assets — and existing answers are either point-scoped tools or hosted ASM products that keep your asset data on someone else's infrastructure. OASM positions itself as the third path: a full ASM stack you run yourself, so the inventory of what you own (and what is accidentally exposed) never leaves your environment. That framing matters for the target audience, because asset inventories are among the most sensitive artifacts a security program produces.
The README organizes ASM into three loops — Discover, Assess, Operate — and OASM implements all three as one integrated system. Discovery enumerates domains, IPs, ports, services, technologies, URLs, and TLS certificates. Assessment runs scanners against those assets and normalizes findings into trackable vulnerabilities with severity, status, and evidence. Operation covers asset grouping, cron-scheduled recurring scans, notification routing, and change auditing. The value proposition is consolidation: instead of stitching together spreadsheets and one-off scanners, a team gets a single searchable workspace with topology graphs, CSV export for reporting, and a job-graph view of recurring workflows.
Architecturally, this is a five-tier design, and the mermaid diagram in the README is unusually informative. Tier one is the web console for daily operations. Tier two is the core API built on NestJS, which holds business logic and orchestration and also hosts the AI Studio agent. Tier three is infrastructure: PostgreSQL with pgvector for persistence and vector search, Redis with BullMQ for job queuing and caching, RustFS object storage for scan artifacts and PDF reports, and a Geo-IP proxy for enrichment. Tier four is the distributed worker fleet, and tier five is connector containers. That separation of API host from worker nodes is a deliberate security decision worth noting, since it means command execution never lands on the API tier.
The connector model is the most operationally interesting part of the design. Scanning tools live as isolated Docker images in a companion repository, oasm-platform/oasm-connectors, each wrapping a tool in a small adapter. OASM reads that catalog as data via task sync-connectors, so a worker resolves the right image and runs the tool on demand inside its own container. The consequence is that adding or upgrading a scanner never requires an OASM release — connectors are decoupled versioned artifacts browsable through a marketplace, then installed from the Tools page in the console. The README shows nmap and rustscan for port scanning, and katana, ffuf, dirsearch, and Arjun for URL and parameter discovery, with more in the truncated list.
Workers themselves are a study in pragmatic distribution. They join over gRPC, pull queued jobs from the BullMQ pipeline, execute each scan in an isolated container, and stream findings back to the core API for normalization. Several workers can share one host, which makes horizontal scaling a matter of spinning up more containers rather than re-architecting the deployment. For teams scanning large estates, that elasticity combined with cron-scheduled asset-group workflows means recurring coverage runs without manual intervention, and on-demand runs remain available per group when you need a targeted sweep.
AI Studio is the headline feature and the one most likely to draw attention. It is a security agent that operates directly on your workspace data and your own workers, using the model you choose — OpenAI, Anthropic, Gemini, DeepSeek, OpenRouter, or a custom endpoint. Ask mode answers questions from inventory and findings; Agent mode turns a goal into a visible to-do plan and works through it step by step. The agent delegates to background subagents running in parallel, each with its own model and limits — the README enumerates planner, researcher, pentest, retester, reporter, and asset analyst roles, plus custom definitions — and every subagent's steps and tool calls are recorded and inspectable in a chat side panel.
The sandboxing description deserves credit for specificity. Shell commands from AI Studio run on a worker node, never on the API host, and on shared workers each session gets its own locked-down container: no host environment or Docker socket access, a read-only filesystem, all capabilities dropped, and CPU/memory limits. The headless browser component drives Chromium on the workspace's worker, can inspect accessibility trees, requests, cookies, storage, and console logs, and confines each session to a separate incognito context whose proxy only permits traffic to the host it opened plus explicitly allowed hosts. For an agent that can execute commands and browse, these are the right blast-radius controls, and the approval model — Manual per command, Plan once for the plan, or Auto — gives operators a dial between oversight and autonomy.
Skills round out the agent: built-in playbooks for reconnaissance, web application testing, network service testing, exploit verification, vulnerability analysis, command execution, browser use, and web research, plus workspace-authored custom skills. Note that these are descriptions of capability for authorized assessment contexts, not something to point at infrastructure you do not own. The MCP server feature extends the story further: any MCP-compatible AI client can connect to your OASM instance with an API key and work with its data from the tools you already use, which is a sensible interoperability move given how fast the agent-client ecosystem is consolidating around MCP.
Integrations cover the operational tail end of ASM. Notifications route to in-app, Slack, Discord, Telegram, and generic webhooks, filtered per integration by event type and severity. Cloud asset sync pulls scheduled inventories from AWS, Cloudflare, and Vercel, keeping the asset list honest against real cloud estates. GitHub and GitLab repositories can be tracked in inventory, and OASM will open an issue for every new vulnerability — a closing of the loop from finding to ticket that most teams currently do by hand. Multi-workspace support isolates teams or projects with owner/member roles, API keys, and an audit log archived as JSONL, which matters for consulting shops running parallel client engagements.
The API surface is versioned REST with interactive Swagger/Scalar docs, backed by a tracked OpenAPI spec that the console client is generated from — spec-first development that keeps client and server in lockstep. Distribution is via Docker (oasm/oasm-api on Docker Hub, per the badge), with a nightly CI workflow visible in the README's build badges, and full documentation hosted at docs.oasm.io covering architecture and the scan lifecycle. There is an active community presence via Discord, LinkedIn, and X, which suggests this is a maintained project with real uptake rather than a one-shot portfolio piece.
For an authorized professional, where does open-asm fit? It is a defender-first tool despite the pentest topics in its GitHub taxonomy: the natural workflow is pointing it at your own or your client's contracted scope, letting discovery enumerate drift between what the CMDB says exists and what is actually exposed, then scheduling connector-based assessment to keep the vulnerability queue current. The pgvector-backed database and AI Studio's inventory-grounded answers make it plausible to ask plain-language questions like which assets run an outdated technology, and get answers sourced from your own scan data rather than a vendor's cloud. That is the strongest argument for the self-hosted model the project is built around.
Caveats worth flagging: the README is long on capability and short on operational hardening guidance beyond the AI sandbox, so teams deploying it should treat the worker fleet as scan infrastructure, network-isolate it, and scope every workflow to owned or authorized targets. The reliance on third-party connectors means scanner behavior and output normalization depend on the companion repo's adapter quality. But as a GPL-licensed, self-contained answer to a problem currently dominated by expensive hosted platforms, open-asm is one of the more credible open-source ASM efforts to appear recently, and the connector-plus-worker architecture means it can grow with the tooling ecosystem rather than behind it.
oasm-platform/open-asm.Educational analysis for authorized security professionals. Use only in controlled, authorized environments.
Related coverage
0 comentários:
Post a Comment
Note: Only a member of this blog may post a comment.