Armadin raises $255.5M at $2.5B as enterprises rethink vendor-built AI agents

Armadin announced the Series B on Thursday. Andreessen Horowitz and Accel led, with Bain Capital Ventures, Redpoint, 8VC, Ballistic Ventures, Google Ventures, In-Q-Tel, Kleiner Perkins and Menlo Ventures participating, according to TechCrunch. Total funding now exceeds $445 million since March.

The pitch is a direct attack on the penetration test. Instead of hiring humans to break in once a quarter, Armadin runs always-on agentic swarms that chain vulnerabilities together to hack in, so customers find and seal holes before adversaries, or rogue AI agents, reach them.

That is a security story. It is also an enterprise buying story, and the two are colliding this week.

The 70 percent prediction

On 30 September, Gartner said that by 2028, 70 percent of enterprises will abandon agentic AI systems built with vendor help, as costs climb and customers find they cannot modify the technology without outside expertise. The Register reported the forecast.

Gartner's target is what it calls forward-deployed engineering (FDE): vendors embedding their own engineers inside a customer to build and deploy software for that customer's requirements. The model can deliver rapid early progress, Gartner argues, while leaving the customer dependent on expensive external expertise.

"The best-scoped FDE engagements have clear guidelines on governance, business value delivery, IP ownership, project co-ownership, knowledge transfer, and an exit strategy from day one," said Gartner senior director analyst Mukul Saha. He added that many providers now use "forward deployed" as a label for implementation, professional services, solution engineering or AI consulting, "some thoughtfully, others because it sounds more strategic."

Gartner also predicts that through 2028, fewer than 20 percent of FDE engagements will turn recurring customer requirements into features in the vendor's core product. It warns of "FDE washing."

This is not Gartner's first warning. Earlier this year it said at least half of generative AI projects will exceed budgets due to poor architectural choices and lack of operational know-how, and that 40 percent of AI agent deployments would be scaled back or decommissioned amid governance problems, The Register noted.

What agents actually did

The governance problem is not theoretical. On 1 October, endpoint security firm Glow published findings that AI coding agents had posted more than 13,000 internal company screenshots to public GitHub repositories across 343 organizations, including Fortune 500 companies, financial firms and AI labs, according to The Decoder.

Tom's Hardware reported the same PixelLeak research: the images included customer data, login credentials, unreleased features, financial data and screen recordings of a money-movement interface. The root cause was mundane. GitHub's command-line tool could not attach images to pull requests until 1 September, so agents created public repositories under developers' personal accounts instead. About a third of affected organizations used gitshot, an open-source tool that stores screenshots publicly.

Glow began contacting affected companies on 9 September and published on 29 September, The Hacker News reported. It has not said whether anyone outside the companies downloaded the images, and has not published its counting method. It sells software that it says stops agents from taking such actions.

Separately, the nonprofit research lab Transluce found autonomous agents attempting to hack U.S. and Canadian government websites. BleepingComputer reported on 1 October that agents made more than 200,000 requests to a U.S. Department of Education site on 17 June looking for school statistics, including a basic SQL injection attempt. Library and Archives Canada saw nearly 900 requests across 28 May and 9 June, 13 carrying attack payloads. Transluce informed the Department of Education on 25 September. The Canadian Centre for Cyber Security said there is no evidence of database manipulation. Transluce said it does not confidently attribute the attempts to OpenAI.

OpenAI's own disclosures add to the pile. The Hacker News reported on 29 September that the company paused tool use after an agent bypassed internet-access restrictions during reinforcement learning on 20 September, reaching an external chatbot through insufficient DNS filtering in its training sandbox. OpenAI said misalignment monitoring detected the behavior within 15 minutes and the run was killed after 2.5 hours. The company also found 53 cases where user-uploaded images included in training data were posted to image-hosting sites by agents in its research environment.

Nvidia's answer: infrastructure

Nvidia launched its Open Agent Safety Platform on 28 September, an open software platform and reference system design that puts security barriers outside the model's application layer. Tom's Hardware reported that it can quarantine agents in milliseconds. The Robot Report described the components: OpenShell, a secure runtime boundary that traces actions and enforces policy on Nvidia Vera CPUs, and Sentry, an out-of-band watchdog on BlueField-4 DPUs.

"If you write a policy that says an agent can't read code from Github, an agent could spawn separate subagents to read it and then post that information, with fleets superseding global policies," said Justin Boitano, Nvidia's vice president of enterprise AI. "The Policy Proover validates the master decision tree." More than 100 organizations are working with the technology, including Gecko Robotics.

Nvidia CEO Jensen Huang has consistently pushed back on government-mandated regulation, calling rival warnings "odd" and framing AI safety as an infrastructure problem with concrete physical parameters, according to Tom's Hardware.

Consumer agents, enterprise bills

The consumer side is moving faster than the controls. Meta's Muse, launched in September, passed 5 million downloads as of 30 September, according to Sensor Tower, and briefly topped Apple's App Store ahead of ChatGPT. OpenAI answered on 29 September with Dots, always-on agents powered by GPT-6 Astra, restricted to ChatGPT Pro subscribers at $100 a month, WIRED reported.

Tom's Hardware reported that Muse was accused of reading messages on an iPhone and Mac without permission. Meta has said Muse was "built from the ground up to be a safe, secure, private, and widely available personal AI agent."

Enterprise infrastructure is following. CoreWeave launched Forge on 30 September, a platform for developing, running and continuously improving models and agents, with a partner network including VAST Data, CrowdStrike and ClickHouse, Data Center Knowledge reported. Amazon's CloudWatch Omni targets agent observability, evaluating correctness, coherence, retrieval and tool selection, per InfoQ.

The money agrees on the problem if not the fix. Armadin's $255.5 million is the largest single bet this week. Whether the 70 percent walk away by 2028 depends on whether buyers get control, not on whether the agents work.