Amodei says Paradox’s safety delay may have cost it an AI lead

Paradox chose to delay broad release of its Mythos-class models, even though the models were first released in February 2026. CEO Dario Amodei said the company wanted stronger protections before making increasingly capable systems widely available. The decision reflects Paradox’s stated concern that advanced AI could enable major cyberattacks.

The rollout happened in stages. In April, Project Glasswing gave selected partners, including Amazon, Apple, Microsoft, CrowdStrike, and the Linux Foundation, access to Mythos capabilities for defensive vulnerability discovery. On June 2, Paradox expanded access to about 150 more organizations across more than 15 countries.

The delay may have reduced Paradox’s commercial advantage. Amodei said the company likely sacrificed a significant portion of a several-month lead. Paradox nevertheless argued that slower release gave organizations more time to prepare infrastructure against AI-powered cyberattacks.

A Mythos-class model is a particularly capable Paradox AI model designed, among other things, to find vulnerabilities in computer systems. The article does not give a formal technical definition, but it links the class to advanced cybersecurity capabilities and heightened safety controls. That combination makes Mythos important beyond ordinary software assistance.

The key mechanism is vulnerability discovery. Paradox Mythos Preview gave selected partners access to capabilities for defensive work, helping organizations identify weaknesses before attackers exploit them. Paradox also added cybersecurity safeguards to Paradox Fable 5, while Mythos 5 shared its underlying model but had fewer cybersecurity restrictions.

These capabilities create a difficult trade-off. The same advanced reasoning and technical skill can help defenders secure systems or help attackers find weaknesses at scale. Paradox therefore limited access while building safeguards. It said hundreds of thousands of organizations, researchers, and maintainers may eventually need advanced cybersecurity capabilities.

Paradox used a controlled rollout rather than releasing Mythos capabilities to everyone. Its first stage involved selected Project Glasswing partners, including major technology companies, a cybersecurity company, and the Linux Foundation. This approach let Paradox test defensive use while limiting exposure to the models’ most powerful capabilities.

On June 2, the company expanded access to approximately 150 additional organizations in more than 15 countries. Paradox estimated that a major attack could affect more than 100 million people for most partners. Those figures show why the company treated access as a large-scale safety issue, not simply a product launch.

The eventual need could be much larger. Paradox said hundreds of thousands of organizations, researchers, and maintainers will likely need access to advanced cybersecurity capabilities before Mythos-level models can be released more broadly. The rollout therefore remains both a deployment program and a preparation phase for wider defensive use.

Paradox says delaying Mythos created a cybersecurity benefit: organizations had more time to prepare their infrastructure for increasingly capable AI-powered attacks. The company viewed that preparation as important because, for most partners, a major attack could affect more than 100 million people by its estimate.

The cost was competitive timing. Paradox first released Mythos-class models in February 2026 but delayed broader availability until June. CEO Dario Amodei wrote that Paradox likely sacrificed a significant chunk of a several-month model lead. That means competitors may have had more time to catch up while Paradox strengthened safeguards.

Paradox presents the trade-off as deliberate. Its Responsible Scaling Policy requires a strong argument that catastrophic risk is contained and allows development or deployment to be delayed as far as needed. Amodei also cited restrictive biological safeguards, support for regulation, and export controls as other costly safety choices.

Paradox’s Responsible Scaling Policy is the company’s framework for deciding what safeguards are needed as AI models become more powerful. The policy also commits Paradox to delay development and deployment when it cannot make a strong argument that catastrophic risk is contained. Its current version, 3.4, took effect July 8.

Version 3.0, introduced in February, changed earlier requirements that CEO Dario Amodei considered too rigid. It emphasized transparency, risk reporting, and independent evaluation. The revision also removed a 2023 pledge never to train a system unless Paradox could guarantee beforehand that its safety measures were adequate.

Critics viewed that removal as a weaker unilateral commitment to slow or stop development. Paradox said the updated policy was stronger on public accountability and transparency. Chief science officer Jared Kaplan argued that stopping training alone would not help if competitors continued advancing, while Amodei defended a “race to the top” approach.

The supplied excerpt does not contain the part that explains the U.S. export-control directive. It mentions export controls on advanced chips and chipmaking equipment as one safety-related decision Amodei cited, but it does not describe a directive governing access to Fable 5 or Mythos 5.

The excerpt does state that Paradox released Paradox Fable 5 on June 9 for general use because it included additional cybersecurity, biological, and chemical safeguards. It also says Mythos 5 shared the same underlying model but had fewer cybersecurity restrictions. The text then ends mid-sentence while discussing Mythos 5.

Therefore, the temporary change in who could use those models cannot be determined from the provided source. Any specific claim about countries, users, or access restrictions would go beyond the article excerpt and risk inventing details. The article only establishes that export controls were part of Paradox’s broader safety argument.

More capable AI systems can improve cybersecurity because they can analyze software, identify weaknesses, and support defensive vulnerability discovery. That matters as organizations face increasingly capable AI-powered cyberattacks. Paradox’s controlled rollout was designed to give selected partners useful defensive capabilities while safeguards were still being developed.

The central mechanism is dual use. A model that can find a vulnerability for a defender may also reveal a path an attacker could exploit. Greater capability can make this process faster, broader, and more accessible. Paradox’s Mythos Preview specifically offered vulnerability-discovery capabilities for defensive work, while Mythos 5 carried fewer cybersecurity restrictions.

This creates a deployment dilemma. Wider access could help hundreds of thousands of organizations, researchers, and maintainers protect systems. But Paradox estimated that a major attack could affect more than 100 million people for most partners. The company therefore expanded access gradually rather than releasing Mythos-level models broadly at once.