Summary: Cognition launched Devin Security Swarm on July 1, 2026 to scan codebases, prove findings in a quarantined sandbox, and open remediation pull requests for human approval. In Cognition’s benchmark over 50 real vulnerabilities, Devin reported 72% recall at $90.23 per run, compared with Claude Security at 68% and $131.87. The product shifts security work toward agent-proposed, human-approved fixes, though the vendor benchmark requires independent skepticism.
Devin Security Swarm went live on July 1, 2026, handing security teams the engineering firepower of Devin, Cognition’s AI software engineer: crawl an entire codebase for vulnerabilities, confirm inside a quarantined sandbox that each finding can genuinely be exploited, then open remediation pull requests that humans review. Cognition’s own framing: it refuses the scanner label — this is an agent product whose whole point is letting security teams push the fixes out themselves. It shares its week with Devin Fusion, Cognition’s move to make frontier performance 60% cheaper, and carries that fight over price into vulnerability remediation.
The Core Problem Beneath the Flood of Alerts
Cognition sizes up the market’s pain in blunt terms: certain teams now receive “10–100x more security findings” than they used to, many of them false positives. Today’s scanners and general-purpose AI utilities fall down on reasoning — business logic, chained attacks, the codebase as a whole — and the most dangerous class of vulnerabilities hides in exactly that terrain. Rules and pattern matching alone just feed an alert queue someone must work through by hand, leaving security staff busier than before. The product’s bet is to supply the missing cross-file reasoning and change the deliverable itself: a list of vulnerabilities already confirmed exploitable.
Inside the Swarm: Find It, Prove It, Fix It
At runtime, a swarm of agents works in parallel, each taking its own segment of the codebase and reasoning across files. Devin stitches the separate findings into a full attack path, replays every path inside a sandbox, and holds the report until exploitability has been demonstrated. The security team receives a verified finding with the attack path laid out and the steps to reproduce it; the patch is the agent’s work too, filed as a remediation PR that merges only after an engineer approves. Find, confirm, fix — three duties that used to sit with different teams now share one workflow, the root difference from tools whose job ends at reporting.
Making Sense of the Benchmark Sheet
The evaluation was built on 50 real-world vulnerabilities, each mapped to the published GitHub Security Advisories, and spans in excess of 14 languages. The scores: Devin Security finished at 72% recall for $90.23 per run; Claude Security posted 68% at $131.87; Codex Security reached 48% at $118.20; Cursor Security trailed at 26% on just $4.60. The misses tell the sharper story: three critical vulnerabilities got past every other tool, caught by Devin alone — a sandbox bypass via PHP template injection, an argument injected through metadata parsing, and a Spring Kafka deserialization surface left too wide. Cognition’s claim: against the closest rival, it surfaces more verified vulnerabilities for 30% less spend. A vendor grading its own homework deserves skepticism, but publishing recall beside the price of each run, on one common yardstick, is roughly the openness this category should expect.
How Setup Works, and the Program for Enterprise Backlogs
Onboarding friction was clearly a design constraint. A scan profile can be spun up from threat-model documents the team already owns — nothing to configure repository by repository, no CI changes. Scheduling accepts daily, weekly, or fully custom intervals, and because later runs touch only changed code, the bill shrinks the longer a team stays. The product is live today, and a number of major companies already put it to work on routine scans. Enterprises carrying deep backlogs got a separate announcement on July 2: a six-week Security Vulnerability Remediation Program in which forward-deployed engineers join the customer’s team, first pointing Devin at the existing CVE backlog, then turning discovery and remediation into an ongoing practice. Snyk, SonarQube, Semgrep — whatever scanners a team already runs can plug in beside it, so the existing stack and the new agent sit side by side, nothing displaced.
Three Takeaways for Engineers and Security Teams
First, remediation now follows an “agent proposes, human approves” rhythm; engineers keep hold of review and merge authority, the design choice that lets a product like this clear an organization’s acceptance bar. Second, proven exploitability has become the new battleground: false positives drain engineering hours, misses leave real exposure, and a sandbox reproduction pulls both down at once — security findings, for the first time, get held to a “reproducible” bar of engineering grade. Third, a tool’s value here now gets computed from its recall and its cost per run together, so teams sizing up rivals can insist on figures on a common baseline and tune out adjectives. With finding volumes climbing at a 10x pace, the coming year’s default arrangement is taking shape: agents produce the triage calls and first-draft patches, while judgment stays a human job.