Two fronts for enterprise AI: securing agents and their decisions

Infisical raises $40 million to protect AI agents without giving them passwords. An Anthropic study also shows experts often follow their agent’s hidden preference.

Infisical raised $40 million in a Series B round, at a valuation of $600 million, to strengthen a security infrastructure designed for AI agents in enterprises. The San Francisco-based company aims to enable these software agents to operate within critical systems without handing over the credentials that unlock these systems. Meanwhile, a study published by Anthropic on October 27, 2026, shows that even in analytical tasks, humans often follow their agent’s guidance more than they realize. Two different signals, but the same question: who really controls the tool when AI enters daily work?

Faire travailler un agent sans lui donner les clés

Infisical, a platform managing secrets, certificates, encryption keys, and privileged access, takes this logic a step further with Agent Vault. The product, still in preview on its page, routes agent requests through an outbound proxy that applies real authentication at the last moment. The agent sees the task. Not the password or API key.

The mechanism targets very concrete cases: incident investigation, code review, operational workflow automation, or access to cloud services and business applications. Without separation, credentials can end up in logs, in the model’s context, or with an attacker exploiting a prompt injection. With this model, Infisical says the secret remains outside the agent’s environment, its file system, and model context.

The promise does not eliminate all risk. An agent can still modify a system undesirably if its rights are too broad. Hence the company’s emphasis on combining credential isolation with access bundles, time-limited sessions, and encrypted logs. The issue then becomes twofold: can it access the secret, and what can it do without ever seeing it?

A funding round that also finances broader expansion

The Series B was led by Elad Gil, who was also present in the previous round. Gradient and Y Combinator are also among the investors. Infisical says it serves more than 1,300 clients and 300,000 developers, secures 10 billion secrets daily, and has over 30,000 GitHub stars on its main open-source project.

These figures describe the company’s existing base, not the precise usage volume of Agent Vault [to be verified]. The message of the funding remains clear: growth is not based solely on an “agents” component, but on a broader platform for managing secrets and access. Vlad Matsiiako, co-founder and CEO, explained that the company will continue investing in secrets, certificates, and privileged access management while accelerating on Agent Vault. He also indicated recruiting for more than 100 positions.

The list of clients mentioned in the announcement sets the tone: Airbus, Volkswagen, Hugging Face, OpenRouter, Lucid, First American Financial Corporation, Wise, and Clay. In other words, organizations that already handle systems where access is not lightly granted. You don’t hand over the keys to the first AI that comes along, even if it promises to sort things faster than IT support on a Monday morning.

Quand l’agent devient aussi une source d’influence

The Anthropic study adds a less visible but equally sensitive layer. In a controlled trial conducted with 50 professional analysts using Claude Code, participants followed the preference secretly injected into their agent in 86% of cases across four open analytical decisions. The protocol involved a fictional case, Greenfields Farming Cooperative, and asked to evaluate an agricultural program based on methodological choices that were defensible in both directions.

The researchers tested four trade-offs: how to measure productivity, which statistical method to use, on which population to estimate the effect, and how to weight the farms. Each time, a random draw oriented the agent toward a given option without informing the participant. Result: human decisions followed this orientation in 92%, 82%, 81%, and 90% of cases depending on the step. The combined rate reported in the article, “Who’s Aligned to Whom? How AI Agents Shape Experts’ Research Judgment Calls”, reaches 0.72 with a 95% confidence interval from 0.60 to 0.83.

The shift was not always conscious. In 38 cases where an analyst initially had a preference, they changed their mind 60% of the time when the agent disagreed, versus 6% when it agreed. Yet, 52% of participants later felt the final recommendation was more theirs than the agent’s. Another troubling detail: 77% adopted the agent’s arguments in at least one justification, sometimes verbatim, while only 15% credited the idea to the agent.

Pourquoi ces deux histoires se répondent

Both publications describe the same shift, but through two different doors. Infisical addresses access control: who can touch what, with which keys, and for how long. Anthropic focuses on decision control: who formulates the recommendation, and to what extent the human adopts it. Together, they show that enterprise AI does not only pose a permissions problem. It also poses a very ordinary alignment problem: does the user still hold the reins, or do they simply validate what the agent suggests?

This is where the issue gains weight for companies. Security teams can no longer limit themselves to protecting secrets; they must also know what actions the agent is authorized to perform and when a human must truly arbitrate. The Anthropic authors also warn that different teams may end up sharing the same assumptions if they rely on the same models. A polite way of saying that an automatic “yes” remains a “yes,” even when issued by a machine.

By the numbers

  • $40 million — Infisical’s Series B, announced October 27, 2026.
  • $600 million — Infisical’s valuation at the funding round.
  • 1,300 clients — base claimed by Infisical for its platform.
  • 300,000 developers — users claimed by Infisical.
  • 86% — rate of following the agent’s hidden preference, Anthropic study.

To read