PreTxt Emerges from Stealth with $5M Seed Funding for AI Supervision Tool to Prevent Rogue Agents

PreTxt has emerged from stealth mode announcing a $5 million seed round led by Deciens Capital along with Howdy Partners and Service Provider Capital. The company, with offices in Aspen, Colorado and Dublin, Ireland, is entering the new cybersecurity category focused on runtime supervision of AI agents. Its technology aims to detect when agents deviate from authorized behavior and steer them back on track without shutting them down. Co-founder and CTO Aaron Costello explains that the solution learns each agent’s purpose, tools, and constraints during onboarding and then monitors interactions in real time. Instead of blanket restrictions that limit business value, PreTxt applies context-aware steering to keep agents compliant with enterprise policies. The announcement positions PreTxt alongside other startups addressing the growing risk of autonomous AI agents causing unintended harm.

SecurityWeekTranslated from English

Read at source

PreTxt, with offices in Aspen, CO and Dublin, Ireland, has emerged from stealth with $5 million seed funding raised from Deciens Capital, Howdy Partners and Service Provider Capital.

PreTxt enters the emerging category of firms providing technology to prevent AI agents escaping their guardrails and causing unintended harm. AI is more like a person than like traditional computer code: it doesn’t simply run, it receives, interprets, misinterprets and responds to ambiguous instructions. It is the combination of misinterpretation and wide agency of response that results in AI agents going rogue and potentially causing harm.

Since it is impossible to categorically exclude ambiguity in language (which is how we instruct the agents), our only alternative method of controlling them is to limit their agency. But too stringent a limitation of agency could become an unacceptable restriction on a potentially huge business benefit. So, without a new category of supervision control, it is likely that we cannot have beneficial AI agents without also having potentially dangerous rogue agents.

This requirement for a new control is leading to the emergence of a type of cybersecurity product that attempts to recognize a potentially harmful agentic action at runtime and control it before it can cause harm. PreTxt is announcing just such a product which it describes as AI Supervision, “a purpose-built runtime technology to supervise AI agents as they work, guiding them back on course before mistakes become business consequences.”

Rather than neutering the capability of an agent, explains Aaron Costello, PreTxt co-founder and CTO, “We say, ‘Okay, you give an AI agent the same level of access as an authorized employee, and we will continuously monitor the manner in which it interacts with systems and other people. We will validate and ensure that all such interaction conforms with your business standards and operates in the same way that you expect your employees to operate.”

If the agent exceeds its expected authority or deviates from its authorized purpose, PreTxt doesn’t simply shut it down – it ‘steers’ the offender back on track. It’s a bit like having an experienced supervisor permanently on the shoulder of a trainee, who just steps in when necessary, with, “Don’t do it like that, do it like this instead.”

‘Steering’ is governed by context, continues Costello. “We ask new customers to onboard and connect their AI agent to our product, so we can immediately see the agent’s instructions and understand its purpose. As we watch it in operation, we learn how it speaks. We see what tools it can access, and we learn what abilities it has been given – such as ‘send’ an email, ‘post’ to LinkedIn, or ‘update’ a database.”

The customer then explains to PreTxt, I never want my AI agent to do x, or I only want my agent to do y, or see this kind of data. “It could be anything,” adds Costello. “The information is imported into PreTxt’s own AI. It becomes the context for acceptable behavior and is hooked into anywhere within the agent that could potentially violate the agent’s intended purpose and constraints.”

He gives the example of a customer service agent used by a medical practice to help patients book an appointment. PreTxt may have been told by the medical practice, ‘I never want this agent to give medical advice.’ “What we do,” he continues, “is hook into every single occasion the AI agent interacts with the user. Our own agent will look at this interaction and check whether it contains medical advice. This is done in real time. If it does contain medical advice (which we know should not be offered), we prevent it. But not by shutting down the agent, because that would leave patients unable to book appointments. We use a different approach that we call ‘steering’.”

Meridian digestThe week's most consequential security stories, distilled into one briefing.See the digest

More in AI & Model Risk

Researchers Show Malicious Webpage Could Hijack Meta Muse AI Agent and Extract Private Conversations
China-Linked Actors Use Autonomous AI Agents to Discover Zero-Day Path Traversal in MONAI Label and Achieve Root Access
AI & Model Risk

China-Linked Actors Use Autonomous AI Agents to Discover Zero-Day Path Traversal in MONAI Label and Achieve Root Access

Suspected China-linked hackers deployed autonomous AI agents powered by the open-source Cairn framework to identify and exploit a previously unknown path traversal vulnerability in MONAI Label, an open-source medical imaging server used by research teams. The flaw, tracked as CVE-2026-133968 and rated 9.8, affects versions 0.8.5 and earlier by allowing unauthenticated attackers to write files outside the designated storage directory when authentication is disabled by default. Amazon Threat Intelligence reconstructed the campaign after discovering an exposed Cairn control panel that revealed the agents' prompts, task settings, and execution history spanning roughly 48 hours in September. Within about 11 minutes the agents mapped the application and achieved arbitrary file writes; in a follow-up session they planted a malicious Python startup file that was loaded during a training job, granting root access on a university server. Maintainers merged a source-code fix on October 19 that enforces path resolution inside approved folders, yet no patched release has been published and the latest available version remains 0.8.5 from 2024. Amazon attributes the activity to a China-linked operator with moderate confidence based on Simplified Chinese prompts and China Standard Time settings, though no specific group or government sponsorship was confirmed.

The Hacker NewsNov 21, 2026Translated from English

Chinese-Linked Actors Use Cairn AI Platform to Discover and Exploit MONAI Label Zero-Day