Anthropic’s Claude Mythos is an AI system capable of discovering and exploiting software vulnerabilities autonomously at a scale no human team can match. What the industry is now absorbing is the next implication since Mythos-class capabilities will not remain isolated to a single provider or a select group of researchers. As advanced vulnerability discovery becomes more accessible across the AI landscape, attackers can now identify and exploit weaknesses faster than ever.
The numbers from Mythos establish what this change looks like. It found over 10,000 high- and critical-severity vulnerabilities across more than 1,000 open-source projects, with a 72% exploit success rate. It developed working exploits for flaws that had survived decades of human security review, autonomously and without human guidance. The Cloud Security Alliance (CSA) published an expedited briefing on this shift, and its conclusions are worth taking seriously.
The instinct will be to respond to each new AI capability as it arrives. But faster vulnerability discovery only changes how fast attackers can get in, not what they’re after once inside. What they’re after has not changed, and the organizations that have already built strong identity security programs will be better positioned than those scrambling to catch up.
The Mythos conversation tends to focus on vulnerability discovery. While the industry debates zero-days, AI agents are already running inside enterprise environments, operating with privileged access and almost no oversight. This does not replace the threats that cause most breaches today, like old, unpatched vulnerabilities and compromised user accounts. It stacks a new privileged identity class on top of them, which is why the CSA brief rates this attack surface as critical.
AI agents hold credentials, execute actions across systems, and operate at a scale and speed no human team can monitor manually. They carry the same risks as any privileged account with broad access, limited oversight, and a significant blast radius if compromised. The difference is that most organizations have not yet applied the same discipline to AI agents as they do to privileged human identities. They cannot fully enumerate what agents are running on their behalf, what those agents can reach, or what they have done.
This is the non-human identity (NHI) problem at machine speed. Every AI agent operating in an enterprise environment is a privileged identity. It needs to be treated as one with scoped credentials, controlled access, and a full session record of every action it takes.
The EU AI Act enforcement deadline arrives in August 2026, and audit capabilities for AI-initiated activities are moving from best practice to legal requirement. Organizations that have not established that capability are running short on time.
But compliance is only the starting point.
Most identity security tools check access at the door and log what happens after. That works if the threat comes from outside. It does not work when an agent starts with legitimate access and drifts into actions it was never meant to take.
Runtime enforcement is the solution.
Every session must be continuously controlled, not just logged
Policy engines need to authorize every action at runtime, evaluating the agent's identity, the resource it is reaching for, and live risk context. When something falls outside policy, the session gets immediately terminated.
Controlling the session only matters if organizations understand what attackers are after once they are inside.
Vulnerabilities provide access to a system. What an attacker does with that access depends on the data they reach, the systems they traverse, and the damage they cause.
The CSA report's own findings make this point directly. The most consequential recent incidents stemmed from credential abuse, social engineering, and supply chain compromise, not zero-days. Faster vulnerability discovery raises the stakes on access, but it does not replace credential abuse as the dominant breach path.
Credentials are still the prize because of the access behind them. Standing access to sensitive systems, broad permissions held by accounts that are never audited, and identities that operate without oversight are the conditions that turn a compromised credential into a significant breach.
The destination has not changed. Mythos just guides attackers there faster.
Identity security is the foundation. Every major AI model provider (Anthropic, OpenAI, Google, NVIDIA) has independently reached the same conclusion. Credentials should not persist in places where they can be reached and exploited. Access should be granted at the moment it is needed, scoped to the task, and revoked when the work is done. That discipline is what limits the blast radius when a breach occurs.
Unmanaged identities and exposed credentials demand a response grounded in accountability, not reaction.
Mythos changes what attackers can do and how fast they can do it. Security leaders do not need to abandon what they know; they need to execute with greater rigor, consistency, and urgency.
The organizations that will lead through this period are those that treat identity security as a strategic investment rather than a compliance obligation.
Full identity visibility, proactive posture management, and continuous authorization controls are the foundation of accountability the Mythos era demands.
Continuously discover every identity and the access it possesses. AI-driven attackers are built to find indirect privilege chains. Continuous discovery across human, machine, and AI identities surfaces the quiet paths their models are built to exploit.
Eliminate standing privileges. Attackers hunt for standing privilege to exploit. If privileges are granted in real time, only when actively needed, a stolen credential has nothing to abuse.
Minimize and broker secrets. Where ephemeral access is possible, eliminate credentials altogether. Where credentials must persist, vault them, rotate them, and broker access so they are only injected at connection time.
Authorize continuously, not just at the door. AI-driven attacks can begin with legitimate access and escalate mid-session. Policy engines must evaluate identity, resource, and risk context throughout the session.
Treat AI agents as first-class privileged identities. Discover every agent, broker its credentials, issue scoped just-in-time access, and govern its activity with the same oversight applied to human administrators.
These priorities are grounded in fundamentals that have not changed, even as the speed and scale of the attacks continue to grow. The threat landscape has evolved. The importance of identity security has not.
To learn more about how Delinea helps organizations build identity security programs ready for the challenges ahead, view our webinar: Mythos: Five Best Practices for Identity Security Leaders