Gizmodo reported on April 17, 2026 that Anthropic’s Mythos AI system is under active consideration for deployment across the entire U.S. federal government — the same federal government whose Trump administration classified Anthropic as a supply chain security risk and barred agencies from using its products roughly 90 days earlier. The reversal is complete. It happened faster than almost anyone predicted, and it involved a lawsuit, a Trump-tied lobbying firm, and a cybersecurity capability so formidable that blocking it started to look like a self-inflicted wound.
The story of how Anthropic moved from federal pariah to potential federal infrastructure in a single quarter is not primarily a story about politics. It is a story about what happens when an AI system becomes too capable to exclude.
The Blacklisting Was Real and It Was Meant to Stick
The Trump administration’s supply chain risk designation was not a bureaucratic footnote. Federal agencies received explicit procurement guidance barring them from Anthropic products — the kind of classification that closes every government contracting door simultaneously. The cited concern involved Anthropic’s corporate structure and its backing from Amazon and Google at a combined valuation exceeding $61 billion. Foreign capital exposure was the stated justification.
The designation gave competitors uncontested access to federal procurement channels Anthropic could not touch. While Anthropic was fighting the designation in court, rival vendors — including OpenAI, which was simultaneously closing a $1 billion content deal with Disney — had no equivalent restriction. The asymmetry was commercially significant and almost certainly intentional.
Anthropic’s Two-Track Response: Courts and Lobbyists
Anthropic did not absorb the designation quietly. The company filed suit and obtained a court injunction blocking enforcement — a meaningful legal outcome that few private technology companies have managed against a federal security classification. The injunction stopped immediate operational damage without resolving the underlying dispute.
On the political track, Anthropic retained Ballard Partners, the Washington lobbying firm founded by Brian Ballard, one of Donald Trump’s most consequential fundraisers and a figure with documented access to key administration officials. The hire carried no ambiguity. A company facing a politically-motivated security designation retaining the most Trump-connected lobby shop in Washington is a signal, not a coincidence.
Both tracks moved in parallel. By the time Mythos‘ capabilities became publicly documented, the political pathway had already been cleared.
The UK AI Safety Institute Put a Number on It: 73%
The policy reversal becomes structurally logical once the underlying evaluation data is examined. The UK AI Safety Institute (AISI) conducted a formal pre-deployment evaluation of Mythos Preview and reported a 73% success rate on expert-level capture-the-flag (CTF) cybersecurity challenges — benchmark tasks calibrated to require the same skills as credentialed offensive security professionals.
For reference, most commercial large language models score below 20% on expert-level CTF benchmarks. Scoring 73% means Mythos is autonomously solving problems that stump working human penetration testers at a rate that has no precedent in published AI evaluation results. The AISI classified the system as warranting enhanced oversight — safety-evaluation language signaling that a model is capable of consequential real-world harm if misdeployed, but also consequential real-world benefit if used defensively.
That framing — dangerous to ignore, useful to control — describes exactly the calculation federal security agencies make about dual-use capabilities.
Project Glasswing: Thousands of Zero-Days, Not Dozens
Anthropic’s internal disclosure about Project Glasswing converted the AISI benchmark into operational reality. Under Glasswing, Mythos Preview has discovered thousands of previously unknown zero-day vulnerabilities in production software — security flaws with no existing patches that could be exploited from the moment of discovery.
Finding a single zero-day vulnerability is a significant achievement in professional security research. Elite private firms and national intelligence services accumulate them in ones and twos over extended campaigns. A system generating thousands at AI inference speeds represents a structural change in how the attack surface of global software is mapped — and an enormous asymmetric advantage for whichever party deploys it first.
The federal government’s interest in Mythos is most clearly explained here. An AI system from a company that has had its own operational security stumbles found thousands of zero-days across external software. The NSA, CISA, and Cyber Command do not need to be persuaded that this capability is relevant to their mission.
JPMorgan Confirmed It Before Washington Did
JPMorgan Chase CEO Jamie Dimon confirmed in April 2026 that the bank is actively testing Mythos for enterprise deployment. JPMorgan operates one of the largest private cybersecurity operations in the financial sector, with an annual technology budget exceeding $15 billion. Dimon’s public confirmation is not a marketing endorsement — it is a due-diligence signal from an institution that cannot absorb a high-profile security failure.
When the largest U.S. bank by assets validates a system for live testing, federal procurement officers assign that validation institutional weight. JPMorgan’s evaluation effectively provided independent credentialing that no amount of Anthropic’s own promotional materials could have generated.
The 90-Day Arc, Mapped
The sequence from blacklist to potential government-wide deployment is worth stating explicitly:
- January 2026: Trump administration designates Anthropic a supply chain risk; federal agencies barred from procurement
- February 2026: Anthropic files suit, obtains injunction blocking enforcement
- February–March 2026: Ballard Partners retained for federal lobbying
- March 2026: UK AISI publishes Mythos Preview evaluation — 73% success on expert-level CTF benchmarks
- March–April 2026: Project Glasswing zero-day disclosures reach public awareness; Jamie Dimon confirms JPMorgan testing
- April 17, 2026: Gizmodo reports Mythos under consideration for government-wide federal deployment
Nothing in this sequence is accidental. The injunction bought time. The lobbying created access. The capability disclosures built the case that excluding Mythos had become a liability, not a precaution.
What This Means Beyond Anthropic
The reversal establishes a replicable template for AI vendors facing politically-motivated regulatory resistance: litigate to stop immediate damage, lobby to shift the political environment, and allow capability demonstrations to build the economic argument for reversal. Movements pushing back against unchecked AI expansion are discovering the same dynamic — capability outpaces political resistance when the capability is genuinely consequential.
MegaOne AI tracks 139+ AI tools across 17 categories. Mythos’ 90-day arc from blacklisted vendor to candidate federal infrastructure is the fastest such reversal in enterprise AI history by a significant margin. It will not be the last — but it has set the precedent for how the next one will unfold.
The outstanding question is not whether Mythos is capable enough for federal deployment — the 73% CTF rate and thousands of Glasswing zero-days settled that. The real question is whether deploying a system with that offensive security profile across federal networks creates more risk than it removes. That is a harder problem than the politics, and nobody in Washington has answered it yet.
