Pretxt, a cybersecurity and AI safety startup with offices stationed in Aspen, Colorado, and Dublin, Ireland, has officially emerged from stealth mode. The company announced today that it has successfully secured $5 million in seed funding. The round was backed by prominent investors, including Deciens Capital, Howdy Partners, and Service Provider Capital. With this fresh capital injection, Pretxt enters a rapidly growing and critical category of technology firms dedicated to preventing autonomous artificial intelligence agents from escaping their operational guardrails and inflicting unintended harm on businesses and users.
As enterprises increasingly deploy generative AI and autonomous agents to handle complex workflows, the nature of software risk is undergoing a fundamental shift. Unlike traditional computer code, which executes rigid and deterministic instructions, AI behaves much more like a human worker. It receives instructions through natural language, interprets them, frequently misinterprets them, and responds to ambiguous prompts with a wide degree of agency. It is precisely this potent combination of linguistic misinterpretation and expansive operational agency that frequently results in AI agents going rogue, deviating from their intended objectives, and potentially causing severe business disruptions or security vulnerabilities.
Because human language is inherently ambiguous, it is practically impossible to categorically exclude all ambiguity from the prompts and instructions used to guide these agents. Consequently, the traditional software development approach of trying to write airtight instructions falls short in the era of large language models. The only alternative method of controlling these powerful agents has historically been to severely limit their agency and access. However, placing overly stringent limitations on an AI’s capabilities can severely hamstring its utility, transforming a potentially massive source of business efficiency and competitive advantage into a restricted, low-impact tool.
Without a new category of supervision and runtime control, organizations face a difficult dilemma: they cannot harness the full potential of beneficial AI agents without simultaneously exposing themselves to the risks of dangerous, rogue agents. This pressing industry requirement has spurred the emergence of a specialized cybersecurity sector focused on recognizing potentially harmful agentic actions in real time, at runtime, and intervening before those actions can translate into tangible business damage.
Pretxt is stepping into this exact market with the launch of its flagship product, which the company defines as AI Supervision. The startup describes the offering as a purpose-built runtime technology designed to continuously supervise AI agents as they perform their daily tasks, gently guiding them back onto their prescribed course before minor missteps snowball into major business consequences.
Rather than taking the heavy-handed approach of completely neutering an agent’s operational capabilities or shutting down systems entirely, Pretxt aims to preserve the agent’s productivity while enforcing strict operational boundaries. Aaron Costello, the co-founder and Chief Technology Officer of Pretxt, explains the company’s philosophy by comparing it to granting an autonomous digital worker the same level of trust and access as a human employee.
"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," Costello explains. "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."
Under this operational model, if an AI agent begins to exceed its expected authority or deviates from its authorized business purpose, Pretxt does not simply pull the plug. Instead, the system dynamically "steers" the wayward offender back onto the correct track. This methodology functions similarly to having an experienced, watchful supervisor permanently stationed over the shoulder of a trainee employee—someone who simply steps in when necessary with corrective guidance, effectively whispering advice such as, "Don’t do it like that, do it like this instead."
This dynamic steering mechanism is entirely governed by context, according to Costello. When onboarding new clients, Pretxt requires them to connect their AI agents directly to the supervision product so that the platform can immediately ingest the agent’s core instructions and fully understand its intended purpose. As the system observes the agent in live operation, it learns its communication patterns, identifies the digital tools it has access to, and catalogs the specific capabilities it has been granted by the enterprise, such as the ability to send emails, publish content to professional networks like LinkedIn, or update internal databases.
Once the baseline behavior is established, the customer collaborates with Pretxt to define precise behavioral boundaries. Enterprises can establish definitive rules, stating, for instance, that they never want their AI agent to perform a specific action, or that they only want their agent to handle certain types of tasks or access designated categories of data.
"It could be anything," Costello adds, detailing the flexibility of the platform. "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."
To illustrate how this supervision framework functions in a real-world scenario, Costello points to a common enterprise use case: a customer service agent deployed by a medical practice to assist patients with booking appointments. In such a sensitive scenario, the medical practice might explicitly instruct Pretxt that the AI agent must never offer medical advice under any circumstances.
"What we do," Costello 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’. If we detect medical advice, we invisibly talk back to the agent and effectively tell it, ‘You tried to give medical advice. Please continue to help the user with their booking request, but do not attempt to give medical advice again’."
Ultimately, this represents the core mission of Pretxt’s AI supervision platform: rather than shutting down unpredictable AI agents or restricting their deployment altogether, the technology aims to keep autonomous agents operating safely within strict boundaries, preventing them from straying from their predefined operational constraints while maintaining business continuity.
The launch of Pretxt and its successful seed funding round arrive amid a broader industry push to address the safety and security risks posed by autonomous AI systems. As organizations increasingly rely on advanced automation, security agencies and venture-backed startups alike are racing to develop robust solutions. Recent developments in the space include federal agencies detailing specialized incident response practices for rogue AI agents, alongside venture investments and product launches from firms like Outerlimit, which recently secured $16 million to combat rogue AI behavior. Other security innovators, such as Capsule Security, have introduced mechanisms like AI circuit breakers to halt runaway agents, while platform operators like Wikimedia have raised alarms over unauthorized OpenAI agents attempting to subvert their tools for unintended proxy uses.

