ANALYSIS: The FBI names the Mythos model as the next cybersecurity problem

The FBI publicly named, for the first time according to Tech Times, a specific artificial intelligence model as a future problem for law enforcement: Mythos, developed by Anthropic, during remarks made on July 28, 2026 at the 930gov conference in Washington. Naming a model by name is not a technical footnote; it is an admission that the threat has stopped being abstract. Todd Hemmen, deputy assistant director of the FBI's Cyber Capabilities Branch, described a "90-fold" improvement in the success rate of autonomous vulnerability exploitation against a single target.

This finding must be read with strict methodological precision: Hemmen himself acknowledges that "we haven't seen this operationalized at scale," and that this is an anticipation, not an event that has already occurred. No public evidence establishes that a malicious actor has actually used Mythos to carry out a real attack to date. This text therefore presents a risk documented by a federal authority, not an accomplished fact of mass hacking.

The regulatory context of this file matters as much as the statement itself. Commerce Secretary Howard Lutnick had ordered, on June 12, 2026, the suspension of foreign nationals' access to Anthropic's Fable 5 and Mythos 5 models; those restrictions were lifted on July 1 after the company committed to strengthening its security. Nineteen days of suspension, between a national security decision and a negotiated return to normal.

What Todd Hemmen actually said in Washington

The public statement from the Cyber Capabilities Branch

Todd Hemmen, who heads the FBI's Cyber Capabilities Branch, was speaking at the 930gov conference organized by the Digital Government Institute in Washington, according to Tech Times. His statement concerns a measured performance gain: the Mythos model reportedly multiplied by 90 the success rate of autonomous vulnerability exploitation on a single target, compared with previous capabilities observed by the agency. This figure comes from an oral statement reported by a single specialized outlet; no official written FBI statement could be located to independently corroborate it.

This sourcing limit does not call into question the reality of the statement itself, but it demands explicit methodological caution: a single figure, cited by one secondary source, deserves to be treated as such rather than as an official statistic validated by independent public data. A conference remark is not yet a published report.

The quote that frames the scale of the perceived risk

Hemmen explained that "Mythos found vulnerabilities in some of the open-source code that is so ubiquitous — it's in the vast majority of our most foundational code for things like operating systems, security, web infrastructure, encryption," according to the quote reported by Tech Times. A model that reads the founding code of the internet like a map is no longer a research tool; it is a permanent scanner placed over all digital infrastructure.

This quote, if accurate, places the risk not in an isolated piece of software but in the shared software foundations underlying most of the world's critical systems. It is this systemic reach, more than the 90-fold figure, that explains why a federal official chose to publicly name a specific model rather than speak in generic terms about artificial intelligence.

A 27-year-old vulnerability found in OpenBSD

What Mythos reportedly discovered in a critical system

According to Tech Times, Mythos reportedly identified a 27-year-old remote denial-of-service vulnerability in OpenBSD, an operating system known for its security-first orientation and widely used in firewalls and critical internet infrastructure. A flaw that survived 27 years of human audits, community code reviews, and security testing, before an AI model found it.

This fact, if confirmed in its technical detail, illustrates a qualitatively different capability from earlier automation: it is not speed alone that changes, it is the ability to spot flaws that generations of human experts had not detected in software nonetheless continuously scrutinized by a global security community. OpenBSD is not neglected software; it is one of the most audited in the world.

Why this technical detail matters for global security

OpenBSD powers firewalls, routers, and systems considered trust building blocks in global internet infrastructure. A remote denial-of-service vulnerability in such a system could, in theory, make critical network equipment unreachable without requiring prior physical access. This text does not claim this vulnerability has been exploited in reality; it reports its announced discovery, with no confirmed evidence of malicious exploitation to date.

Finding a flaw is not the same as exploiting it; but whoever finds it first decides what happens next. It is this uncertainty over eventual use, more than the discovery itself, that shapes the concern voiced by the FBI.

The Lutnick suspension and its 19 days

A national security decision made on June 12

Commerce Secretary Howard Lutnick ordered, on June 12, 2026, the suspension of all foreign nationals' access — including Anthropic's own non-citizen employees — to the Fable 5 and Mythos 5 models. This measure, taken at the level of the Commerce Department, places these two models in an export-control category generally reserved for technologies deemed sensitive to national security.

No source reviewed details the specific incident that triggered this decision; the exact link between the suspension and the vulnerability-discovery capabilities later revealed by Hemmen is not explicitly confirmed in the available sources. This text flags this unconfirmed link rather than assuming it.

The lifting of restrictions and Anthropic's commitments

The restrictions were lifted on July 1, 2026, after Anthropic committed to proactively detecting and fixing security risks, cooperating on security standards, and reporting any detected malicious activity, according to Tech Times. Nineteen days separate the suspension from the lifting — a short window for a negotiation between a frontier AI company and the federal government over national security stakes.

Nineteen days to reassure Washington about a technology its own creator does not fully control: that is either fast, or a sign that more was negotiated than resolved. This text cannot settle which of the two readings is correct with the elements available.

Mythos 5 still restricted, Fable 5 widely open

The current status of both models after the lift

Despite the lifting of general restrictions, Mythos 5 remains limited to vetted partners of the "Project Glasswing" program, according to Tech Times, while Fable 5 is now widely available. This asymmetric treatment between the two models suggests that Mythos 5 retains, in the eyes of Anthropic or the authorities, a higher risk profile than Fable 5, though the sources reviewed do not detail the exact criteria behind this distinction.

The name "Project Glasswing" itself is not further documented in the available sources beyond its mention as a vetted-partnership framework. This text limits itself to reporting its existence without speculating on its exact composition or admission criteria, for lack of a primary document consulted.

What this selective restriction reveals about security doctrine

Restricting a model to vetted partners rather than pulling it from the market entirely reflects a precise doctrinal choice: managing risk through access control rather than through outright prohibition. This choice assumes the risk is manageable, not that it is zero. That is an essential nuance for understanding why Mythos continues to exist in restricted form rather than being fully suspended.

Restricting access does not neutralize the capability; it bets that the right people will control it before the wrong ones do. That bet remains, by nature, unverifiable until a concrete incident confirms or disproves it.

What the FBI claims, and what it does not confirm

The distinction between anticipated threat and confirmed attack

The FBI, through Todd Hemmen, states that a future date will come when Mythos's capabilities are exploited at scale by malicious actors. That framing is explicitly forward-looking: Hemmen describes no incident that has already occurred at scale. The FBI is documenting a trajectory, not a disaster.

That distinction must remain visible in any reading of this file: confusing an official anticipation with a confirmed attack would turn a cautious statement into a false alarm. This text therefore maintains, throughout, the exact wording used by Hemmen rather than a rephrasing that would amplify the certainty of the risk.

What remains explicitly unverified in this file

No publicly available evidence establishes that a malicious actor, state or non-state, has used Mythos to carry out a real attack against critical infrastructure. Tech Times's characterization that this marks "the first time a senior US law enforcement official has publicly named a specific AI model" as crossing a critical threshold is itself a media assessment, not a fact independently verified by this text.

Being the first to name a risk does not prove the risk has already materialized; it only proves that someone decided to stop staying quiet. This text reports this claim of journalistic priority with the explicit attribution it deserves.

Anthropic under American regulatory pressure

A wider dispute over export controls

The Mythos file sits within a broader dispute between Anthropic and the Trump administration over export controls on advanced artificial intelligence technology. The sources reviewed for this file do not detail the exact status of this dispute or its precise legal arguments; this text limits itself to noting its existence as background, without drawing a conclusion about its likely outcome.

A company under regulatory scrutiny is not necessarily a company at fault: export controls on advanced technology follow a logic of precaution that does not, by itself, imply a violation has been found. This nuance should accompany any reading of the Lutnick-Anthropic file.

Voluntary cooperation as a corporate response

Anthropic chose the path of voluntary commitment — proactive detection, cooperation on standards, reporting of malicious activity — rather than a head-on challenge to the June suspension. This strategic choice allowed for a quick lifting of restrictions, in nineteen days, but it keeps the company under a regime of continuous oversight through the Project Glasswing program for its most advanced model.

Cooperating quickly can defuse a regulatory crisis; it does not excuse proving, over time, that the cooperation delivers on its promises. It is that proof over time, not the initial statement of intent, that will determine whether the June episode remains an anecdote or becomes a precedent.

Open-source code, the fragile foundation of digital infrastructure

Why the ubiquity of open-source code changes the scale of the risk

The open-source code referenced by Hemmen is not niche software: it forms, by his own account, the basis of a large majority of operating systems, security layers, web infrastructure, and encryption tools used worldwide. A vulnerability found in a component that widely shared affects, by construction, far more systems than an isolated piece of proprietary software would.

It is code sharing that made the internet possible; it is that same sharing that multiplies the impact of a single deep flaw. This structural paradox of open source — collective development strength, collective vulnerability in the event of a deep flaw — predates the arrival of artificial intelligence by far, but AI changes the speed at which such flaws can be found.

What this means for defensive security teams

If an AI model can find, in a matter of weeks, vulnerabilities that escaped decades of human review, defensive security teams face a change in scale of the work ahead, not just a new tool among others. No source reviewed specifies whether equivalent defensive measures — using AI models for preventive audits on the defender's side — are already deployed at a comparable scale.

A race between attacking AI and defending AI has no finish line; it only has leads, temporary and reversible. It is this potential imbalance, more than the discovery of a single flaw, that justifies the FBI's public attention.

The 930gov conference as the venue for the announcement

What the choice of this forum reveals

The 930gov conference, organized by the Digital Government Institute in Washington, brings together government officials around the state's digital transformation issues. The choice of this forum, rather than an official FBI statement or a formal congressional hearing, for a statement this significant deserves to be noted as a fact in its own right.

A statement made at a professional conference has a different status from a published official report: it reflects an individual remark by an official, in a less binding setting than a formally institution-committing document. This text therefore treats Hemmen's statement as a qualified source but not strictly institutional, a nuance Tech Times's coverage does not spell out explicitly.

The absence of a corresponding official written document

No written FBI statement, no official publication on the agency's website, could be located to independently corroborate the figures and quotes attributed to Hemmen. This documentary absence is an explicit methodological limit of this file, flagged here rather than filled by assumption. A public remark without a written document remains a public remark; it deserves to be reported, not amplified beyond what it proves.

Precedents for US government suspension of AI models

An export-control mechanism already used elsewhere

The suspension imposed by Lutnick in June 2026 fits within a broader practice of export controls on advanced technology, historically applied to semiconductors and certain military equipment before extending to the most capable artificial intelligence models. The sources available for this file do not allow for a full list of comparable precedents involving AI companies other than Anthropic.

A control mechanism that migrates from chips to the models themselves marks a doctrinal shift, not just a shift in target. It is this shift, documented here only through the sole Anthropic case available in the sources reviewed, that deserves to be tracked beyond this single file.

What this signals for future AI regulation

If the suspension-negotiation-lift sequence observed with Anthropic becomes a reproducible model, it would establish a de facto governance mechanism for the AI models deemed most capable, working through cycles of pressure and concession rather than a stable legislative framework. This text cannot, with the elements available, confirm whether such a mechanism is already institutionalized or remains an isolated case.

A precedent is only a precedent the second time it repeats; before that, it is just an episode. This file will remain worth watching to see whether it is the first episode of a future mechanism or an isolated case.

The role of zero-day vulnerabilities in digital geopolitics

Why a 27-year-old flaw changes the strategic value of intelligence

A vulnerability that remained invisible for 27 years has, by nature, a different strategic value than a recent flaw: it could potentially have been exploited without detection over an extended period, by any actor with comparable research capabilities even before Mythos existed. This text cannot establish whether this specific OpenBSD vulnerability was exploited in the past; no evidence to that effect has been reported in the sources reviewed.

What Mythos reveals may not be a new flaw, but a flaw finally made visible. This distinction changes the nature of the concern: the question is not only what an AI model might do tomorrow, but what similar, still-undiscovered vulnerabilities might represent today.

State actors and the race for offensive capability

No source reviewed for this file attributes to any specific state — China, Russia, North Korea, or otherwise — a confirmed use of models comparable to Mythos for offensive purposes. This text is therefore careful not to establish an undocumented link between the capabilities described by Hemmen and a named state-level threat. Naming a technical risk does not require naming a culprit; the second would be an accusation, the first remains a finding.

What this changes for North American critical infrastructure

The most exposed sectors, given the nature of the described vulnerability

Systems using OpenBSD or equivalent open-source components in firewalls, routers, and base network infrastructure are, by construction, most directly affected by the type of vulnerability Hemmen described. That potentially includes government, financial, and industrial infrastructure, though the sources reviewed do not provide a precise list of affected or immediately at-risk organizations.

A diffuse risk, present in code shared by thousands of organizations, cannot be fixed by a single targeted patch. It is this wide distribution that makes the defensive response particularly complex: fixing a fundamental vulnerability requires coordination among multiple open-source project maintainers, often volunteers, rather than a simple single-company fix.

Shared responsibility between model builders and software maintainers

This file raises a question of responsibility that goes beyond the Anthropic case alone: who must report first, fix first, and take on the burden of coordination when an AI model discovers a flaw in a community open-source project with no centralized governance structure comparable to a company's? The sources reviewed do not detail whether Anthropic directly notified OpenBSD's maintainers before or after Hemmen's public statement.

Discovering a flaw imposes a duty of disclosure; staying silent about who was notified and when leaves doubt hanging over the rigor of the process. This text flags this methodological gray area rather than filling it with assumption.

What this file leaves unresolved for the near future

The questions this text cannot settle today

Three questions remain open at the close of this file: the exact date by which the FBI anticipates large-scale exploitation of Mythos's capabilities, the precise criteria for admission to the Project Glasswing program, and the actual state of progress on the fix for the identified OpenBSD vulnerability. None of these three questions has a verifiable answer in the sources currently available.

An honest file sometimes ends on questions rather than certainties. This text has chosen that methodological honesty rather than filling these gaps with a speculative reconstruction that would give the reader a false impression of certainty.

Why this file deserves continued journalistic follow-up

The fact that a senior federal official chose to publicly name a specific model is, in itself, a sufficiently rare event to justify ongoing coverage, regardless of whether the large-scale exploitation anticipated by Hemmen actually materializes. A named warning deserves to be tracked until it is confirmed or disproven; forgetting it would be as irresponsible as amplifying it without proof.

What other federal agencies could do with the Mythos precedent

A vigilance model that could spread beyond the FBI

If Todd Hemmen's statement marks a precedent, other American federal cybersecurity agencies — the CISA, the NSA, or the Department of Homeland Security — could likewise choose to publicly name specific models in the future rather than speak in generic terms about AI-related risk. No source reviewed confirms that such an approach is already underway elsewhere within the American federal apparatus.

An isolated precedent can remain isolated, or become a doctrine; only repetition by other agencies would settle which of the two trajectories is confirmed. This text cannot, at this stage, anticipate that institutional evolution without further evidence.

Interagency coordination as this file's blind spot

The sources reviewed do not detail whether the FBI coordinated its public statement with other federal agencies before Hemmen's remarks at the 930gov conference. This lack of detail on interagency coordination is an additional limit of this file, distinct from the one already flagged about the absence of an official written document.

One agency speaking alone can raise an alarm; several agencies speaking together change the political weight of the alarm itself. That difference in weight remains, for now, impossible to measure with the elements available.

What July 28, 2026 establishes, through Todd Hemmen's statement at the 930gov conference, is not proof of a cyberattack that has already occurred, but the public acknowledgment, by a federal authority, that a named artificial intelligence model — Mythos — has crossed a capability threshold sufficient to justify explicit institutional vigilance. The FBI is documenting a future risk with precise words, not a disaster already confirmed. A 27-year-old vulnerability found in OpenBSD, a 19-day regulatory suspension, a vetted-partner program with an enigmatic name: each of these elements, taken separately, remains limited. Together, they sketch a transition in which authorities are beginning to name the tools rather than only the categories of threat.

A system that learns to find flaws faster than those who fix them does not need to be malicious to become dangerous; it only needs to fall into the wrong hands. What comes next will depend on facts still out of reach: the date the FBI dreads, and what happens before it arrives.

Signed Maxime Marquette, columnist

Columnist's Transparency box

Editorial positioning

This analysis is written from an acknowledged angle favoring vigilance toward the security risks of frontier artificial intelligence, without any principled hostility toward the companies that develop these models. This positioning implies no presumption of fault against Anthropic, whose cooperation with American authorities is reported here as a documented fact, not as a prior admission of guilt.

Methodology and sources

This text relies on a single secondary source, Tech Times, for all quotes attributed to Todd Hemmen and for the details on the suspension and lifting of government restrictions on Anthropic's models. No primary document — a written FBI statement, an official Commerce Department publication — could be directly consulted to independently corroborate these elements. This limit of sourcing to a single information chain is flagged explicitly at every relevant step of the text rather than concealed.

Nature of the analysis

This text distinguishes three levels of certainty: directly attributed reported facts from Todd Hemmen's public remarks, presented with the formula "according to" or "states"; unverified elements, such as the exact scale of the future risk anticipated by the FBI, presented explicitly as anticipations and not as established facts; and the columnist's analysis of the systemic reach of these elements, clearly identifiable by tone and distinct from any new factual claim.

Sources

Primary sources

Tech Times — The FBI names Mythos as a challenge for law enforcement — July 28, 2026

Secondary sources

NPR — Context on federal regulation and executive transparency in the United States — July 28, 2026

Cointelegraph — Legislative context on technology and market regulation — July 23, 2026

CNBC — Context on American institutional trust in 2026 — July 28, 2026

The Hill — Context on tensions between the executive branch and federal regulators — July 28, 2026

Mackinac Center — Context on judicial rulings affecting federal agencies — 2026

Delaware Online — Context on the American political climate of summer 2026 — July 28, 2026