Trent AI raises $13m to build security agents for AI agents

On 7 April Trent AI, a London company that builds security agents for teams developing software and AI agents, came out of stealth with a $13m (£9.8m) seed round. LocalGlobe and Cambridge Innovation Capital led it. The angels include Joaquin Quiñonero Candela, a member of technical staff at OpenAI, Databricks distinguished engineer Ippokratis Pandis, AWS director Avinash Bhat and Tony Jebara, formerly vice-president of engineering and head of AI/ML at Spotify, according to the company’s announcement.

The product runs four sets of agents. One scans code, infrastructure, dependencies, agents and runtime behavior; one separates real risk from noise and weighs its business impact; one patches vulnerabilities, opens pull requests and adjusts configurations; and one tracks security posture over time and benchmarks it against standards. Trent AI says each cycle makes its agents better at judging the systems they protect.

“Organizations are deploying AI agents and autonomous workflows faster than their security can adapt, and most development teams using these agents and workflows have no security framework designed for their systems,” said Thereska, co-founder and CEO, in the announcement. The company’s argument is that the costliest flaws are set at the design stage, before a scanner has any code to read, and that agents now build and change systems too quickly for one-off human reviews to keep up.

The seed round pays to test that argument outside a small circle of early users. So far the full platform is with design partners, and the launch announcement pointed visitors to a waitlist. The $13m gives Trent AI time to show that a security layer which reasons about design can be sold to teams with no security engineers of their own.

An engineer’s case against security dashboards

Thereska has more than 20 years’ experience building secure, scalable and autonomous systems. He was an early engineer at Confluent, a principal engineer at AWS and a distinguished engineer at Alcion, which Veeam acquired, and earned his PhD at Carnegie Mellon University, according to his company biography.

In a post published with the funding news, he wrote that at large companies, even those with strong security teams, “security often showed up after the code was written”. At Alcion, a few engineers built a threat model of the backup system before any code existed, which he says shaped a safer architecture but relied on individual effort rather than a system. He also described scanning tools that produced “400+ findings at times”, most of them false positives or low-signal noise that engineers ignored.

Smaller companies had a different version of the problem. “Most teams simply couldn’t afford dedicated security engineers, but the stakes were just as high,” he wrote.

A bench from machine learning research

His co-founders come from machine learning research. Lawrence, the chief scientist, is the DeepMind professor of machine learning at the University of Cambridge and a former director of machine learning at Amazon. Dai, the CTO, was a machine learning scientist at AWS and a senior research manager at Spotify. The leadership team also includes Christoph Bartenstein, chief product officer, who spent 10 years at AWS and four at Microsoft, according to the company.

Companies House records show Trent AI Limited was incorporated on 28 January 2025. Thereska and Dai have been directors since that day and each holds more than 25% but not more than 50% of the shares. Lawrence was also a director from incorporation until 19 June 2025.

Two launches before the round

Trent AI names Canopy, Commscentre, ML@Cam, Qbeast and Weblogic among the design partners with early access. It says they have reported immediate visibility into their security posture, a security audit report and fast identification of vulnerabilities. It has not published customer numbers or revenue.

Two products went out in the 15 days before the funding announcement. On 23 March the company released a security assessment skill for OpenClaw, the open-source runtime for autonomous agents, which checks gateway security, tool permissions, Model Context Protocol (MCP) servers and file permissions, and groups findings by severity. On 2 April came a Security Advisor for Lovable, the app-building platform, which reviews a project as it changes and, once the user approves a plan, applies the fixes in Lovable through an MCP connection. Bartenstein, who wrote both announcements, pitched the Lovable tool at builders who “aren’t security engineers”.

The company cited Deloitte’s 2026 State of AI in the Enterprise report, which it says found that 74% of companies plan to deploy agentic AI within two years and 21% report a mature model for governing autonomous agents. Deloitte surveyed 3,235 leaders in 24 countries in August and September 2025.

The investors framed the round in the same terms. “Agent adoption is outpacing enterprise security readiness,” said Ian Lane, partner at Cambridge Innovation Capital. Saul Klein, co-founder and executive chairman of Phoenix Court, the home of LocalGlobe, said Trent AI combined “deep academic expertise with real-world experience building large-scale systems”. According to Tech.eu, the company said the money will go on developing the agents, hiring engineers and growing the base of design partners and customers.

A company incorporated 14 months ago, with five named design partners and two integrations, now has the money to find out whether teams building agents will let other agents review their designs before the code is written.