depthfirst
AI security platform that detects, triages, and remediates software vulnerabilities.
depthfirst is an applied AI lab building what it calls General Security Intelligence, a platform of custom AI agents that detect, triage and remediate software vulnerabilities across code, dependencies, secrets, infrastructure and runtime.
It was founded in 2024 by Qasim Mithani, former head of infrastructure security at Databricks, and Andrea Michi, who spent nearly seven years building AI at Google DeepMind, with a founding team drawn from DeepMind, Databricks and Faire. The San Francisco company has raised $120 million in total.
An $80 million Series B led by Meritech in March 2026, at a $580 million valuation, came less than ninety days after its $40 million Series A led by Accel, with angels including Jeff Dean.
The thesis is that winning in the AI era of security requires security specific models, not general language models bent to the task. depthfirst builds its own models, trained through reinforcement learning in security environments: dfs-mini1 for smart contract vulnerabilities, and dfs-large1, post trained on GLM 5.2, in preview for large enterprise repositories. On top of the models sit six named products that work as a continuous loop. A Code Scanner traces business logic and data flows, and a Security Reviewer checks every human and agent code change before merge.
A Dependency Firewall blocks malicious packages, and Secrets and Supply Chain agents follow, the latter tracing reachability through transitive dependencies in every scan over a shared Component Graph of the customer's repositories. An Agentic Pentester validates findings against the running application and replays the attack after a fix is merged, so a finding closes only when exploitation fails. Fixes arrive as pull requests an engineer reviews and merges, and every human and agent action is logged. Access is role based by repository, environment or organization, with customer held encryption keys and a SOC 2 audit.
The company reports many times more true vulnerabilities than static analysis, roughly eighty five percent fewer false positives, and about eighty percent of its fix suggestions accepted and merged. Those are its own figures. Customers it names include Lovable, Supabase, Moveworks, AngelList, ClickUp and incident.io. It does not publish its hosting model or pricing, and its documentation sits behind the customer login. For an engineering organization shipping AI generated code faster than it can secure it, depthfirst validates and fixes findings and does not just flag them. Teams wanting a traditional scanner or a bring your own model approach will find a model owning platform instead.
Vendor details
Canonical URL
https://depthfirst.com
Category
Security / SOC agent
Subcategory
Application security vuln triage and remediation
Funding status
Independent, founded in 2024 by Qasim Mithani, former head of infrastructure security at Databricks, and Andrea Michi, formerly of Google DeepMind, with a team from DeepMind, Databricks, and Faire. Raised 120 million dollars in total, including an 80 million dollar Series B led by Meritech Capital in March 2026 at a 580 million dollar valuation, following a 40 million dollar Series A led by Accel. Angel investors include Jeff Dean and Kirsten Green.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Integrates directly into developer workflows, reviewing pull requests and delivering ready to merge fixes, and connects across code, dependencies, secrets, infrastructure, and runtime in one continuous system. The agentic pentester can validate findings against a running application with or without code access.
In practice
Your team ships AI generated code faster than security can review it. depthfirst's Security Reviewer checks every human and agent change in the pull request and proposes ready to merge fixes, so security keeps pace with development.
Scanners bury your team in false positives. depthfirst's Agentic Pentester validates each finding against your running app with real attack paths, so only exploitable risk reaches the queue, cutting noise sharply.
A vibe coded internal app pulls in a malicious dependency. depthfirst's Dependency Firewall blocks malicious packages before they land, letting you use AI coding tools across the company safely.
Sources & related URLs
Agentic Index coverage score
10.5 / 14 capabilities · 75%
| Integrations & Tool Calling | Full |
|---|---|
|
depthfirst connects to GitHub, GitLab, Jira, Linear, Docker, AWS, Google Cloud, Azure, Terraform and JFrog, and its remediation agent opens a pull request against the customer's repository for every confirmed vulnerability. Sourcedepthfirst.com/platformread 2026-09-28 |
|
| Workflow Orchestration | Full |
|
Several agents hand work to each other across one loop. The code scanner and supply chain agent find issues, the agentic pentester validates exploitability, the remediation agent opens the fix, and the pentester replays the attack after merge, with a finding closed only when exploitation fails. There is no stated workflow builder the customer configures. Sourcedepthfirst.com/platformread 2026-09-28 |
|
| Knowledge Grounding & RAG | Full |
|
The platform maps the customer's repositories into a Component Graph of data flows, cross service relationships and every dependency. Supply chain scanning runs on that same graph, a shared layer beyond the context of a single run, though how long it persists between scans is not stated. Agents show the call chain behind each finding, and product docs sit behind the app login. Sourcedepthfirst.com/platform and depthfirst.com/supply-chainread 2026-10-25 |
|
| Human Oversight & Guardrails | Partial |
|
Policy set once applies to every scan, fix and agent action, a customer controlled constraint. Fixes arrive as pull requests an engineer reviews and merges, but that review gate belongs to the customer's code host, and there is no stated depthfirst approval step before an agent action commits. Sourcedepthfirst.comread 2026-09-28 |
|
| Security, Identity & Governance | Full |
|
Role based access is scoped by repository, environment or organization, and customers hold their own encryption keys. The SOC 2 attestation is independently audited and has a trust center at trust.depthfirst.com, but its type is not given. Sourcedepthfirst.comread 2026-09-28 |
|
| Observability & Auditability | Full |
|
Every human and agent action is logged, immutable and traceable, showing who did what, when and why, and each finding carries the full call chain from application code to the vulnerable function. Sourcedepthfirst.comread 2026-09-28 |
|
| Memory & State Persistence | Partial |
|
The platform improves over time by learning from developer feedback on its recommendations, which carries state across scans. There is no stated memory store, scope or lifetime. Sourcedepthfirst.com/platformread 2026-09-28 |
|
| Deployment & Data Residency | Not documented |
|
depthfirst publishes no hosting model, region or self hosted option, and customer held keys are an encryption control, not a deployment option. The product docs at docs.depthfirst.com sit behind the app login. Sourcetrust.depthfirst.comread 2026-09-28 |
|
| Prebuilt Agents, Templates & Packs | Partial |
|
There are six named products, Code Scanner, Dependency Firewall, Agentic Pentester, Security Reviewer for pull requests, Secrets and Sensitive Data, and Supply Chain, each doing its own job. depthfirst presents them together as one platform that covers the full cycle and does not say whether each can be adopted on its own. Sourcedepthfirst.comread 2026-10-06 |
|
| Triggers & Channel Coverage | Full |
|
Work starts on events with no one asking. Every pull request is reviewed before merge, the pentester replays attacks after every merge, reachability runs in every SCA scan, and the dependency firewall watches packages in real time. Sourcedepthfirst.comread 2026-09-28 |
|
| Model Flexibility & Routing | Partial |
|
The agents run depthfirst's own security models (dfs-mini1, and dfs-large1 in preview, post trained on GLM 5.2), and the company writes that its harness routes each task to the best models across providers. That routing stays with depthfirst, with no provider list the customer selects from and no bring your own model path. Sourcedepthfirst.com/post/why-defenders-cant-bet-on-one-modelread 2026-10-25 |
|
| APIs, SDKs & MCP Extensibility | Full |
|
An API queries findings, triggers scans and integrates depthfirst into the customer's own tooling, and its reference sits behind the customer login at docs.depthfirst.com. Sourcedepthfirst.com/platformread 2026-09-28 |
|
| Testing, Debugging & Optimization | Full |
|
The pentester replays the same attack after every fix is merged, and a vulnerability counts as resolved only when exploitation fails in the running application, a quality gate on the agent's own output. Recommendations also improve on developer feedback, an optimization loop after deployment. The dfbench model benchmark is depthfirst's own publication, not a customer surface. Sourcedepthfirst.com/platformread 2026-09-28 |
|
| Browser & Computer Use | Partial |
|
Agentic Pentesting, launched 30 September 2026 through demo request, has agents authenticate and navigate the application's UI and APIs. The pages name no browser, and no clicking or form filling. Sourcedepthfirst.com/agentic-pentestingread 2026-10-25 |
|
The Agentic Index coverage score grades every vendor Full, Partial or Not documented against the same 14 buyer facing capabilities, from public evidence only. Each capability links to how all vendors in the index score on it. How this evidence is graded
Product updates
depthfirst launched Agentic Pentesting, autonomous agents that log in to a running application and attack it through its interfaces and APIs, on demand or from a CI pipeline. With source code connected, they use earlier findings and the threat model to pick attack paths, confirm whether scanner or bug bounty findings can really be exploited, and replay the attack after a fix to prove it worked.
Bears on: Agent capability
View sourcedepthfirst makes transitive dependency reachability available to customers as part of every software composition analysis scan. Its agent follows execution paths through intermediate packages and exposes the chain of calls leading from application code to a vulnerable function. The analysis can follow invocation mechanisms that static call graphs miss, including command line startup and framework callbacks.
Bears on: Agent capability
View sourceDepthFirst introduced dfs-large1, a proprietary security model built on GLM 5.2 and post-trained using reinforcement learning within its security agent harness. The model is now available in preview for vulnerability discovery and validation across large enterprise repositories.
Bears on: Agent capability
View sourcePricing
Not published. Sold through sales after a demo.
No billing unit is published.
What is public
No public rate. depthfirst publishes no list pricing, and commercial terms are set through sales.
Billing mechanics
Sold through sales after a demo. depthfirst does not publish its billing unit, and whether the six products are bundled or sold separately is not stated.
Cost watchouts
Whether the Dependency Firewall, Security Reviewer and Agentic Pentester are bundled or priced separately is not published, and neither is how pricing scales with codebase size and scan frequency.
Variable cost rationale
No billing unit is published, so what is metered is not known. Nothing published says whether the price tracks repositories, developers or the products enabled. Every price is quoted through sales, and no cap, minimum or commitment is published that would limit the bill.
Additional watchouts
No rate is published, so the cost cannot be set against current AppSec tooling or security headcount before a quote. Which of the Security Reviewer, Dependency Firewall and Agentic Pentester modules a base contract includes is not published.
Sales call required
Yes, required for paid access
Free / trial
None published. A demo is available on request.
Lowest paid plan
Not published
Key ambiguities
No entry rate or per developer figure is published. All pricing is quoted through sales.
Related vendors
- 7AI — Agentic SOC from the Cybereason founders: domain specialized agents…
- Abnormal AI — Behavioral AI email security with three named autonomous agents: a…
- AirMDR — AI native MDR whose AI analyst investigates every alert and writes…
- Airrived — Agentic OS that unifies SOC, GRC, IAM, vulnerability management, IT,…
- Andesite — Human-AI SOC platform where customers build their own agents and…
- Anvilogic — Agentic security operations platform that works on top of existing…
Alternatives to depthfirst
The closest documented capability profiles to depthfirst among security and SOC agents tracked by Agentic Index, ordered by similarity on the same 14 point evidence the rankings use. No vendor pays for placement.
- Snyk11.5 / 14Adds documented Deployment & Data Residency
- Capsule Security9.0 / 14Fuller documented coverage on Human Oversight & Guardrails
- Repello AI8.0 / 14A lighter documented profile than depthfirst
- RunSybil8.0 / 14A lighter documented profile than depthfirst
- Arctic Wolf9.5 / 14Adds documented Deployment & Data Residency
- Magnitude8.5 / 14Fuller documented coverage on Prebuilt Agents, Templates & Packs
Similarity is computed from each vendor's Agentic Index coverage score evidence, axis by axis, not from the totals. How this evidence is graded