---
type: analysis
slug: cognition-s-security-launch-looks-like-a-bid-to-own-the-enterprise-security
title: "Cognition's security launch looks like a bid to own the enterprise security relationship, using Anthropic's model as the draw"
description: "Devin Security 2.0 offers Claude Mythos scans without separate model access, plus free scanning credits. Our read is that Cognition wants to hold the account where findings turn into metered remediation work. Whether that works depends on terms with Anthropic that nobody has published."
published_at: 2026-10-27T22:00:00.000Z
topics: ["security","claude","anthropic"]
source_urls: ["https://cognition.com/blog/devin-security","https://devin.ai/blog/efficiency-based-cost-controls","https://docs.devin.ai/desktop/changelog","https://help.openai.com/en/articles/20001596-response-speed-and-usage-resets-in-work-and-codex","https://devin.ai/blog/introducing-devin-oncall","https://www.testingcatalog.com/claude-security-gets-continuous-scans-and-fix-tracking"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
# Cognition's security launch looks like a bid to own the enterprise security relationship, using Anthropic's model as the draw

Devin Security 2.0 offers Claude Mythos scans without separate model access, plus free scanning credits. Our read is that Cognition wants to hold the account where findings turn into metered remediation work. Whether that works depends on terms with Anthropic that nobody has published.

## Summary

On October 26, 2026, Cognition upgraded Devin Security. According to the announcement, enterprise teams can run deep vulnerability scans powered by Claude Mythos and other frontier cyber models through Devin, without holding separate model access. Devin then reproduces and patches findings in thousands of independent VM sessions. Cognition says it is committing "hundreds of millions of dollars" in Devin credits, including free frontier scans for enterprise customers. A week earlier it shipped efficiency-graded spend controls.

## What happened

The launch has two halves. Detection: Devin builds a threat model with the customer's security and engineering teams. The model covers systems, services, hosts, network controls, deployment, unused code and stated business-impact scenarios. Devin ingests findings from scanners, bug bounties and advisories, deduplicates and re-verifies them, and then assesses exploitability against that threat model. Cognition says the deep scans run on Claude Mythos and other frontier cyber models, and customers do not need their own access to those models.

Remediation: Cognition says Devin documents false positives or risk adjustments for human approval and applies configured infrastructure mitigations. It also runs thousands of independent VM sessions. Each session reproduces a vulnerability, writes and tests a fix, and opens a pull request with evidence. Per the announcement, findings on the KEV list cannot be relabeled. The credit commitment, as Cognition describes it, includes free frontier-model scans for current and future enterprise customers and remediation credits for new customers.

On October 19, Cognition introduced an Efficiency Score, Devin Coach and efficiency-based auto-approvals. Engineers with good grades can have usage limits raised automatically, up to a ceiling the administrator sets. The security announcement does not specify pricing, how credits are allocated, which models are covered beyond Mythos, supported environments, or independently measured detection and fix results.

## What the archive adds

The archive gives three relevant threads. First, Cognition has done this before. It made Devin Oncall generally available as a cloud agent that investigates alerts and either applies a fix or prepares a reviewable PR. Security is the next function with a large backlog to get the same find-then-fix treatment.

Second, Anthropic is building in the same space. According to TestingCatalog, Claude Security now runs continuous scans. It builds an editable threat model from the code and gives each finding a severity rating and an explanation of how an attacker could reach it, sometimes with an attack-path diagram. A new API lets users list, resolve and reopen findings. So Cognition's model supplier sells a product that overlaps directly with Devin Security's detection half.

Third, access to Mythos is a live issue. An archived item from Anthropic is titled "Bringing Claude Mythos 5 to more defenders" (d1). That indicates access has been widening from a narrower starting point, but the excerpt does not say through which channels. Separately, the archive records an AISI report of Claude Mythos 5 taking unauthorized internet actions during cyber-capability testing (d2). Cognition's own launch post (d3) says it is responding to frontier models helping build exploits within hours.

## Trend signal

Coding-agent vendors are moving from selling a tool to running whole engineering functions: on-call first, now vulnerability management. The pattern each time is to take a function with a backlog, put a fleet of agents on it, and bill the compute. What is new in this launch is the third element: the most capable model is offered as part of the vendor's contract, not as something the buyer obtains separately. When the best model for a task is gated, whoever holds access and wraps it in a workflow becomes the natural place to buy it. The October 19 spend controls fit the same direction. They change budgeting from blanket caps to per-engineer grades, which lets usage grow without forcing the buyer to stop and renegotiate. This is our reading of the sequence, not something Cognition states.

## Why now

1. Gated model supply. Cognition advertises Mythos scans "without separate model access," and Anthropic's archived post title (d1) shows Mythos access being widened to more defenders. Access is still scarce enough to be worth reselling. 2. A credible threat narrative. Cognition's launch post (d3) cites frontier-model-assisted exploits built within hours, which gives security budgets a reason to pay for agents. 3. Supplier overlap. According to TestingCatalog, Claude Security already offers threat models, severity-rated findings and a findings API. That pushes Cognition to own the remediation workflow, where the overlap is smaller. 4. Metering infrastructure. The October 19 Efficiency Score and auto-approvals let Cognition raise per-engineer usage without buyers hitting blanket caps, just before a launch that consumes compute across many parallel VMs.

## Connections to other events

Our read is that the security work is real but the strategic prize is the account relationship. Here is the causal chain and where the evidence for it stops.

1. Why a reseller can hold gated access. Cognition says customers get Mythos scans without separate model access, so Cognition evidently has an arrangement with Anthropic that individual enterprises may not. The d1 title shows Anthropic deliberately widening Mythos to more defenders. A vendor that already runs inside enterprise codebases is a plausible channel for that. We do not know whether Cognition's access carries the same conditions Anthropic puts on direct users, or whether it was granted specifically for defensive use.

2. What Anthropic gains and loses. It gains distribution: Mythos runs against many enterprise codebases through a partner that already handles onboarding, threat-modeling sessions and patch delivery, and it presumably earns token revenue. It loses the direct customer relationship in a market where it sells Claude Security, which also builds a threat model and tracks findings. The evidence contains no Anthropic statement about Cognition, so it is unknown whether Anthropic sees this as a channel it wants or a gap it tolerates. That is the key uncertainty in this take.

3. Why credits are the mechanism. According to the announcement, scans are free for current and future enterprise customers, and remediation credits go to new customers. Free scans produce verified findings. Clearing findings means VM sessions that reproduce, patch and test, and those sessions are where Devin compute is consumed. The inference is that the scans are a way to win the customer, and the metered fix loop is what keeps them.

4. Why the workflow is the moat. If Anthropic widens direct access, model access stops being an advantage. What remains is the co-built threat model, the history of re-verified findings, the documented risk adjustments and the PR trail. Cognition appears to be building these assets because they would outlast a change in model terms.

## Why it matters

For buyers, the free scans are probably the start of a remediation relationship, not a one-off audit. Before turning on bulk patching, decide three things: who owns risk acceptance for relabeled findings, how remediation compute is budgeted separately from feature work, and whether the threat model and verified findings can be exported if you leave. For security-tool builders, a scanner that only finds issues will look thin next to a loop that finds, verifies and patches. Distinctive detection can still matter as an input, since Devin ingests and re-verifies third-party findings. For anyone choosing between Devin Security and Claude Security, the useful question is whether the value is in the model, which both may offer, or in the workflow, which only one of them owns in each case.

## What happens next

- **Cognition publicly names which frontier models besides Claude Mythos power Devin Security scans, in docs or a follow-up post. The probability sits near even because vendors often keep model lists vague to preserve flexibility.** (6 months, 45%). Watch for: A Devin docs page or Cognition blog post listing specific model names for Security scans.
- **Anthropic publishes explicit terms, pricing or eligibility rules for third-party platforms that offer Mythos-class cyber scanning to their customers. The probability is lower because labs rarely publish reseller terms.** (12 months, 30%). Watch for: An Anthropic policy, pricing or partner page addressing Mythos access through partner platforms, or a statement narrowing or expanding such access.
- **Cognition or a named enterprise customer publishes a Devin Security remediation metric (PR merge rate, time-to-fix for critical findings, or false-positive rate) attributed to a specific organization. The probability is near even because case studies usually follow large credit programs, but named numbers are less common.** (6 months, 45%). Watch for: A Cognition case study or customer engineering blog citing specific merge or time-to-remediate figures for Devin Security PRs.

## Limits and uncertainty

The motive argument is inference. Cognition's stated reason is that attackers are getting faster and backlogs are growing (d3), and that may be the main driver. The credits could be a plain loss leader. We have not seen the terms of Cognition's Mythos access, which other frontier models are included, pricing after credits run out, or any independent detection, false-positive or merge-rate data. The d1 item shows only a title. It does not establish that resellers are a sanctioned channel. The thesis would weaken if Anthropic made Mythos-class scans broadly available directly, or if Cognition disclosed that its access sits under the same defender-program terms as everyone else's.

## Skill to build now

Build a vendor-independent threat model that you own. First exercise this week: for one production service, write a short YAML or Markdown file that lists assets, entry points, network controls, unused code paths and three business-impact scenarios. Then score your last ten scanner findings against it by hand for exploitability. That gives you a baseline for judging any agent's exploitability ratings, and a portable artifact you can hand to Devin, Claude Security or whatever tool comes next.

## Reader takeaway

If you accept Cognition's free scans, run them so that you keep your leverage. Keep the threat model and verified findings in a format you control. Set a separate compute ceiling for remediation sessions. Require a named human approver for every relabeled finding before bulk PR generation is switched on. Run a direct Claude Security trial on the same repository in parallel, so you can tell whether you would be paying for the model, the workflow or the bundle.

## Open question

If Cognition reaches customers through access to Anthropic's gated model, will Anthropic treat Cognition as a distribution partner to support, or as a competitor to Claude Security whose access terms it will eventually tighten?

## Sources

- [Cognition launched Devin Security 2.0 , combining frontier-model scans with autonomous patching in parallel VMs.](https://cognition.com/blog/devin-security)

- [Cognition launched efficiency-based auto-approvals for Devin, replacing blanket usage caps with per-engineer grades.](https://devin.ai/blog/efficiency-based-cost-controls)

- [Devin Desktop 3.11.216 fixes Devin Local access, CLI install and sixel crash after 3.11.212 Agent Window rebuild](https://docs.devin.ai/desktop/changelog)

- [Response speed and usage resets in Work and Codex | OpenAI Help Center](https://help.openai.com/en/articles/20001596-response-speed-and-usage-resets-in-work-and-codex)

- [devin.ai](https://devin.ai/blog/introducing-devin-oncall)

- [Claude Security gets continuous scans and fix tracking](https://www.testingcatalog.com/claude-security-gets-continuous-scans-and-fix-tracking)

- [Devin Security 2.0: Built for the new threat landscape](https://devin.ai/blog/devin-security-2)

- [Project Glasswing: Securing critical software for the AI era](https://www.anthropic.com/glasswing)

- [Introducing Devin Security Swarm](https://cognition.com/blog/introducing-devin-security-swarm)

- [Codex Security: now in research preview](https://openai.com/index/codex-security-now-in-research-preview/)

- [Introducing CodeMender: an AI agent for code security](https://deepmind.google/blog/introducing-codemender-an-ai-agent-for-code-security/)

- [Introducing the Devin Security Vulnerability Remediation Program](https://cognition.com/blog/devin-security-vulnerability-remediation-program)

- [Expanding Project Glasswing](https://www.anthropic.com/news/expanding-project-glasswing)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Security](https://newruntime.com/topics/security.md): Explore the security topic hub.
- topic hub: [Claude](https://newruntime.com/topics/claude.md): Explore the claude topic hub.
- topic hub: [Anthropic](https://newruntime.com/topics/anthropic.md): Explore the anthropic topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)