CodeWall raises $4 million for AI agents meant to take over from the annual pen test
CodeWall, a London startup, launched with $4 million for AI agents that attack enterprises, pitched as a replacement for the once- or twice-yearly pen test. A buyer weighing the annual test against it has the company's own runs against consulting firms to go on.
The Product Desk
Key figures: CodeWall launched with $4 million in new funding for autonomous AI agents that attack enterprises. Tests spaced evenly across a year are 365 days apart if commissioned once a year, and about 182 days apart if commissioned twice a year.
| Figure | Value | Note | Claim |
|---|---|---|---|
| New funding for CodeWall's AI attack agents | $4 million | 1 | |
| Tests once a year | 365 days apart | 18 | |
| Twice a year, about | 182 days apart | 18 |
What happened
- Several agents, each built for a different kind of job, work on a target at once, and the company says a weak spot in one application can become another agent's route into connected systems.
- At McKinsey, an agent that started without credentials could read and write to the production database underlying Lilli, the firm's AI platform, inside two hours.
- Research released over the following weeks turned up exposed databases at Bain and Boston Consulting Group.
- McKinsey told The Register, through a spokesperson, that a forensics investigation by a third party had turned up no evidence of CodeWall or any other unauthorized party accessing client data.
Why it matters
- constraint Verification and remediation workflows are still being added, so a buyer today is judging the attack half of the product while the fix half is still roadmap.
- decision Because the agents keep to the customer's authorized scope, the scope document sets the real coverage, and any system left out keeps the old gap between tests.
- exposure The urgency in the pitch rests on Sysdig's own account of an autonomous ransomware operation, which CodeWall uses to argue defenders need the same automation. Teams on a yearly test cycle are the ones that argument is aimed at.
Take a security lead at a bank whose pen test finished in March and whose developers ship a new customer-facing application in June. Spaced evenly, tests commissioned once or twice a year [7] fall about 365 or 182 days apart [18]. CodeWall's pitch is aimed at that interval: its agents go back in whenever new systems come online or a fix is applied [5].
Founder and CEO Paul Price calls the ongoing testing "a permanent capability" [8]. The record is narrower than that. Earlier this year CodeWall turned its agents on the Big Three management consulting firms [19], and each firm fixed its flaws within days of being told [10]. That is a time-to-fix figure. Retention and usage depth are not reported, and neither is pricing.
Price said enterprises "shouldn't have to wait for a real attacker to find out whether their defenses hold." [13] Paul Wilkes, a partner at Osney Capital, which led the round [14], said that in his view having an attacker working for an enterprise will one day be "as fundamental as having defenses in place." [15] His sentence is in the future tense.
We'd keep the annual test, run the agents beside it, and retire the calendar test one system at a time. The cost is paying for overlapping coverage until the agents have matched the human report on a given system. Two facts sort each system. The first is how often it changes, because re-testing after new systems and fixes is where the agents differ from a calendar test [5]. The second is whether the agents can reach it today, and CodeWall says most of the new money will go toward letting them reach more parts of an enterprise's environment [16].
- Changes often, within reach: move first, and retire the annual test for that system once the agents' findings match or beat the human report. - Changes often, out of reach: the highest exposure. Keep the annual test and add a manual retest after each release. - Rarely changes, within reach: run the agents as a second check, with no rush to drop the annual test. - Rarely changes, out of reach: leave it on the annual cycle.
What to watch
- Whether CodeWall publishes pricing, customer counts or retention figures after the launch.
- Whether the verification and remediation workflows ship, and what fix times customers report with them.
- Whether any customer drops its annual pen test in favor of the service.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+30
- Incentives70
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
London-based offensive security startup CodeWall Ltd. launched with $4 million in new funding to give enterprises such as banks autonomous AI agents that attack them the way an outside hacker would.
- [2]
CodeWall begins each engagement with no credentials or inside knowledge of the target.
- [3]
Several agents, each built for a different kind of job, work on a target at the same time; according to the company, whatever one finds is logged to a common investigation graph, so a weak spot in a single application can become another agent's route into connected systems and data.
- [4]
Customers see each completed attack path step by step, with the evidence attached.
- [5]
The agents go back in whenever new systems come online or a fix is applied.
- [6]
The agents keep to whatever scope the customer has authorized.
- [7]
CodeWall is pitching the service as a replacement for the penetration tests most enterprises commission once or twice a year.
- [8]
Founder and Chief Executive Paul Price calls the ongoing testing "a permanent capability."
- [9]
Research published in the weeks that followed found exposed databases at Boston Consulting Group and Bain & Co.
- [10]
Each of the firms fixed its flaws within days of being told.
- [11]
Over the summer, cloud security company Sysdig Inc. documented what it said was the first ransomware operation run from start to finish by an autonomous AI agent.
- [12]
CodeWall argues that defenders now need the kind of automation attackers already have.
- [13]
Price said enterprises "shouldn't have to wait for a real attacker to find out whether their defenses hold."
- [14]
Osney Capital led the pre-seed round, with InReach Ventures, Antler Global Ltd. and angel investors participating; Paul Wilkes is a partner at Osney Capital.
- [15]
Wilkes said that in his view, having an attacker working for an enterprise will one day be "as fundamental as having defenses in place."
- [16]
Most of the money will go toward letting the agents reach more parts of an enterprise's environment.
- [17]
The platform is also getting workflows for verifying and remediating what the agents find.
- [18]
Tests spaced evenly across a year are about 182 days apart if commissioned twice a year and 365 days apart if commissioned once a year.
- [19]
Earlier this year CodeWall turned its agents on the Big Three management consulting firms.
- [20]
A McKinsey spokesperson told The Register that a third-party forensics investigation found no evidence that client data had been accessed by CodeWall or any other unauthorized party.
- [21]
At McKinsey & Co., an agent with no credentials gained read and write access to the production database behind the firm's Lilli AI platform within two hours.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- siliconangle.comOffensive security startup CodeWall launches with $4M for AI agents that attack enterprises
1 article · October 14, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- AI agent securityFollow
- Offensive securityFollow
- Cybersecurity venture fundingFollow
- Autonomous penetration testingFollow