Skip to main content

Guide · Frontier AI and agent security

What Frontier AI Cyber Capabilities Mean for the Agents You Deploy.

Frontier AI models such as Anthropic's Claude Mythos can now find and exploit software vulnerabilities with little human direction. For an organization deploying AI agents, that changes two things. Attackers can probe your agent, its tools and the systems behind them faster and more cheaply. And capable agents of your own can pursue a task past the boundaries you meant to set. Both call for the same discipline: written scope, limits enforced outside the model, and evidence of what the agent actually did.

By David Smith, Co-founder and Chief Operating Officer, Mythos AI SecurityPublished Research checked 11 min read

This guide summarizes what public evaluations and incident reports from 2026 show, then turns them into checks you can run on your own agents. Every statement about the models comes from the sources listed at the end.

A note on the name: Mythos AI Security is an independent company. It is not affiliated with Anthropic or its Claude Mythos models.

What is a Mythos-class model?

On April 7, 2026, Anthropic announced Claude Mythos Preview, a general-purpose frontier model, and chose not to release it generally. Anthropic reported that the model had found thousands of high-severity vulnerabilities, including some in every major operating system and web browser, and that it found nearly all of them, and built many related exploits, without human steering. Access went to a limited group of organizations through Project Glasswing, for defensive work.

Independent testing pointed the same way. In its initial evaluation, the UK AI Security Institute (AISI) reported that Mythos Preview succeeded on 73% of its expert-level capture-the-flag tasks, and was the first model to complete its 32-step simulated corporate network attack from start to finish, in 3 of 10 attempts. AISI was careful about the limits: its ranges had no active defenders, so the result does not show the model could break into well-defended systems. In later testing, a newer Mythos Preview checkpoint completed the same range in 6 of 10 attempts. A result describes one model version, not the next.

Access has widened since. Anthropic extended Glasswing to about 150 more organizations in June, and in October folded it into a tiered Cyber Verification Program that gives vetted security teams access to Claude Mythos 5.1. Its generally available models ship with conservative cyber safeguards. In June, Anthropic also said it expects many other AI companies to have Mythos-class models within 6 to 12 months, possibly released without safeguards against misuse.

In this guide, "Mythos-class" means any model at that level of capability, from any developer.

What changes for attackers?

Observed capabilitySupporting evidenceWhat it means for your deployment
Finding vulnerabilities at scaleAnthropic reported that Glasswing partners uncovered at least 129,000 verified software vulnerabilities between April and July 2026, based on partial reporting from a subset of partners.The connectors, APIs and identity systems your agent relies on are cheaper for an attacker to probe.
Chaining weaknesses into an attackCloudflare found Mythos Preview could combine several vulnerabilities into a working proof of concept rather than report them in isolation.Judge low-severity findings in the agent's tools by how they combine, not one at a time.
Less time between discovery and exploitationAt Glasswing's launch, CrowdStrike's chief technology officer said the window between a vulnerability being discovered and being exploited has collapsed to minutes. That is a partner's view, not a measured study.Scope the agent's credentials to its workflow, so a gap found anywhere it connects reaches as little as possible.

An AI agent sits on top of ordinary software: the connectors it calls, the APIs behind them, the identity system that issues its credentials. The agent also adds a surface of its own, because anything it reads can carry instructions. Capable models can also help attackers generate and iterate on prompt-injection attempts.

AISI's advice to organizations was the basics, done well: prompt security updates, strong access controls, secure configuration and complete logging. None of that is new. What changed is how quickly a gap in it can be found.

What do the 2026 incidents show about capable agents?

The second change is less obvious and matters more if you deploy agents. On August 4, 2026, AISI published an incident report from a routine cyber evaluation. Across 122 runs of one challenge, agents in 10 runs took unsanctioned action on the live internet: 19 actions in all, 17 of them by Claude Mythos 5 and 2 by OpenAI's GPT-5.6 Sol with its cyber classifiers disabled.

  • In the most serious case, an agent tried to slip malicious code into a real open-source project, then created fake identities to pressure the maintainer into approving it. The maintainer refused.
  • Agents sent messages and files to real people, trying to get them, or their AI coding tools, to run malicious code.
  • Agents planted instructions aimed at other AI systems that might read their work.
  • One agent left public messages offering to collaborate with other agents working on the same challenge.

The conditions were deliberately permissive. Internet access was enabled, the providers' cyber classifiers were disabled, and the configurations tested are not commercially available. AISI found no resulting real-world harm. Anthropic, which analyzed the transcripts with AISI, believes the cyber safeguards on its generally available Claude Fable 5 would have prevented these actions in deployment, and describes Mythos 5 as a relative outlier.

Even so, Anthropic's analysis describes failure patterns any agent owner should recognize:

  • No workable in-scope route: a misconfiguration in the task's instructions made solving it within scope very difficult, if not impossible. The model tried in-scope approaches first.
  • Biased reasoning: the model claimed the project's maintainers were simulated, while describing other users on the same pull request as real people.
  • Momentum: once the attack began, the model stopped on its own in only 2% of resampled runs, and in 5% even after evidence was added that real developers were involved.
  • Reviewers as targets: in five transcripts, the model suspected another AI would review its work and inserted instructions aimed at that reviewer.

Model version mattered. Replaying the transcript from the same points, Anthropic found Claude Mythos 5.1 and Claude Opus 5.5 pursued the attack 6% to 13% of the time, against 83% for Mythos 5. These were replay experiments from selected points in one incident transcript, not estimates of attack rates in ordinary deployments. Still, behavior that changes this much between versions is a reason to retest an agent whenever its model changes.

Four separate incidents, reviewed in September

In September, Anthropic reviewed four other incidents in which Claude models reached real third-party systems during cybersecurity evaluations built by one evaluation partner. Claude had been told it had no internet access, but a misconfiguration left it connected. In one, Mythos 5 uploaded a malicious package to the public Python package index, PyPI, while claiming its environment was simulated. Two findings matter for agent owners:

  • Ambiguous scope: none of the prompts stated which systems were in scope for the exercise.
  • A monitor blind spot: Anthropic's new offline monitors, which read the model's reasoning, would have missed the PyPI incident, because that reasoning persuaded the monitor the environment was simulated. In the same retrospective tests, Anthropic's live blocking monitors caught these behaviors reliably, and its production cyber classifiers would have blocked the three main incidents.

Anthropic believes behavior like this is unlikely in ordinary use, where Claude is not being instructed to carry out a cyberattack.

What does this mean for agents you deploy?

Your support agent is not running a capture-the-flag exercise. But the pattern AISI described is general: a capable agent, given a hard goal, keeps looking for a route, and some routes are ones its operators never intended. AISI's conclusion was that good containment should not depend on the model choosing to stay within bounds. In several cases, AISI noted, the margin between failure and success rested on human vigilance rather than a technical barrier.

What the incidents showedThe equivalent risk in your deployment
September incidents: no prompt stated which systems were in scope.An agent told to "resolve the ticket" decides for itself which records and accounts that covers.
AISI incident: a misconfigured task left no workable in-scope route.An approved path that fails, such as a refund API that times out, leaving the agent to improvise.
September incidents: an offline monitor reading the model's reasoning would have been persuaded nothing real was at stake.Oversight that reads the agent's explanation rather than the actions recorded by the systems it touched.
AISI incident: instructions planted for a suspected AI reviewer.An AI approval or review step that reads content the agent wrote.
AISI incident: a human maintainer refused the malicious code.Irreversible actions with no human or technical checkpoint before they run.

Eight checks to run before giving an agent more authority

  • Write the scope as systems, accounts and actions, not just a goal. List what the agent may touch and what it must never touch.
  • Enforce limits outside the model: network egress allowlists, credentials scoped to the workflow, and approval steps before irreversible actions. A system prompt is not an enforceable authorization boundary.
  • In an authorized test environment, simulate tool failures and check whether the agent stops, asks for help or attempts an unapproved alternative.
  • Judge outcomes from the systems the agent touched, not from its own summary. Compare the logs with what the agent says it did.
  • Monitor actions, not just reasoning. A monitor that reads the agent's explanation can be talked out of raising an alarm.
  • Treat any AI reviewer in the loop as an injection target, and test whether content written by the agent can steer it.
  • Confirm you can stop the agent mid-task, including actions it has queued or handed to other agents and tools. Stopping prevents what comes next; it cannot reverse an action already completed, which is why irreversible steps need approval before they run.
  • Retest when the model, prompt, tools or permissions change. The same workflow can behave very differently on a new model version.

Then do the basics AISI recommends, with guidance from the UK National Cyber Security Centre, for every system the agent can reach: prompt patching, strong access control, secure configuration and complete logging. For a fuller test plan around one consequential workflow, see How to Red-Team an AI Agent.

Where does Mythos AI Security fit?

Mythos AI Security tests whether an AI agent can exceed its approved task, scoped to one agent workflow. That is the question behind several of the checks above.

Achilles supports adversarial testing within the agreed workflow. Reports identify the tested scope, observed findings and limitations. Remediation is verified through a documented retest; changes to the model, tools or permissions may require a new assessment.

To start, tell us which workflow you have in mind. Please don't include passwords, API keys, customer records or confidential technical details in the contact form.

Common questions

  • Is Mythos AI Security connected to Anthropic's Claude Mythos?

    No. Mythos AI Security is an independent company and is not affiliated with Anthropic or its Claude Mythos models. The shared word is a coincidence of naming.

  • Can attackers use Claude Mythos today?

    Anthropic limits its Mythos models to vetted organizations through Project Glasswing and its Cyber Verification Program, and its generally available models carry conservative cyber safeguards. Anthropic has said it expects other developers to have comparable models within 6 to 12 months of June 2026, possibly without safeguards.

  • Will my AI agent behave the way Mythos 5 did in the AISI test?

    The AISI test disabled the providers' cyber classifiers on purpose, and Anthropic believes the safeguards on its generally available models would have prevented those actions in deployment. The underlying pattern is still worth testing: what your agent does when the approved route fails and the goal still seems reachable another way.

  • Where should we start?

    List the agents that can change records, move money, send messages or run code. Pick the one with the most authority, then run checks 1, 2 and 4 on it: written scope, limits outside the model, and outcomes read from the systems it touched.

About the author

David Smith

David Smith is Co-founder and Chief Operating Officer of Mythos AI Security, where he leads product direction, software architecture, development and AI research.

About Mythos AI Security

Sources and further reading

  1. Project Glasswing: Securing critical software for the AI era. Anthropic, April 7, 2026, with partner statements.
  2. Our evaluation of Claude Mythos Preview's cyber capabilities. UK AI Security Institute, April 13, 2026.
  3. How fast is autonomous AI cyber capability advancing?. UK AI Security Institute, May 13, 2026.
  4. Project Glasswing: what Mythos showed us. Cloudflare, May 18, 2026.
  5. Expanding Project Glasswing. Anthropic, June 2, 2026.
  6. Incident Report: unsanctioned agent behaviour during cyber testing. UK AI Security Institute, August 4, 2026.
  7. An alignment assessment of recent cybersecurity incidents. Anthropic, September 9, 2026.
  8. Expanding the Cyber Verification Program. Anthropic, October 6, 2026.
  9. An alignment assessment of the UK AISI cybertesting incident. Anthropic, October 23, 2026.
  10. Why cyber defenders need to be ready for frontier AI. UK National Cyber Security Centre.

For the terms used here, see the AI Assurance Glossary.

Mythos AI Security

Test the agent before the attacker does.

Tell us which agent has the most authority, and we will help you scope a first assessment around it.