GLASSWING
“The day defense became the moat.”
A machine found a flaw that outlived four presidents.
In the winter of 2026, a machine did something no human had done in twenty-seven years. It read a piece of software called OpenBSD — the operating system beloved by paranoid sysadmins and the people who run firewalls for nuclear facilities — and found, sitting inside its TCP stack, a flaw so elemental that two carefully shaped packets could knock the whole thing over.
The bug had been there since the Clinton administration. It had survived the dot-com crash, the rise of the smartphone, four presidents, and what one engineer estimated was probably five million automated fuzzing runs. The machine found it on its lunch break.
The compute bill came to less than fifty dollars.
THE 90× GENERATION
“Not a scaling-law improvement. A phase transition.”
VULNERABILITY CHAINING
- 01Two minor Linux kernel bugs braided into a full root compromise in an afternoon.
- 02Four medium-severity Firefox flaws composed into a Just-In-Time heap spray — escaping the renderer and the OS sandbox in a single chained maneuver.
- 03An FFmpeg parsing quirk plus a codec invariant, chained to remote code execution.
“In the world Mythos inhabits, there is no such thing as a low-severity bug.”
THE SaaSPOCALYPSE
Two trillion dollars of enterprise value, exhaled.
“Goldman compared the trajectory of enterprise software stocks to American newspapers in the early 2000s — meaning the decline does not end until earnings have collapsed.”— Goldman Sachs, research note
The quiet meeting the public never saw.
On a Tuesday in early spring, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell summoned the chief executives of America's largest banks to an emergency meeting.
The agenda did not concern interest rates or the yield curve. It concerned a single terrifying question — and the consensus answer, according to two people in the room, was "not enough."
- Brian MoynihanBank of America
- Jane FraserCitigroup
- David SolomonGoldman Sachs
- Ted PickMorgan Stanley
- Charlie ScharfWells Fargo
- Jamie DimonJPMorgan Chasevia letter
CAN OUR DECADES-OLD MAINFRAMES SURVIVE A WEEKEND WITH MYTHOS?
THE TWELVE
A gated, curated, almost feudal coalition of twelve launch partners — permitted to use Mythos for exclusively defensive purposes.
Conspicuously absent: the U.S. Department of Defense.
THE PENTAGON FEUD
- FEB 2026
Hegseth calls Anthropic
The Secretary of Defense tells Anthropic its model will be made available to the military "for all lawful purposes," including two the Constitution explicitly forbids: autonomous lethality and mass domestic surveillance. Anthropic says no.
- 3 MAR 2026
Supply-chain designation
Invoking 10 U.S.C. § 3252 — a statute normally reserved for hostile-nation telecom equipment — the DoD formally designates Anthropic a "Supply-Chain Risk to National Security." First time ever applied to an American firm.
- 9 MAR 2026
Anthropic sues
Filings in both the Northern District of California and the D.C. Circuit. First Amendment retaliation, Due Process, and APA claims.
- LATE MAR 2026
Preliminary injunction
A federal judge in San Francisco grants the injunction, noting the government's behavior looks like "classic First Amendment retaliation."
- ONGOING
The case continues
Fortune 500 CISOs, European data regulators, Swiss private banks, and Middle-Eastern sovereign wealth funds have chosen the vendor the Pentagon blacklisted.
THREE STRATEGIES
ANTHROPIC
CARTEL
OPENAI
DEMOCRACY
GOOGLE DEEPMIND
EMPIRE
WHAT THE CISO MUST DO NOW
Three pillars of the agentic SOC.
PREEMPTIVE INVALIDATION
Patches must run at machine speed.
Finding-to-fix is easier for an AI than finding-to-exploit. Triage committees are dead. Remediation must be continuous, automated, and instantaneous.
DYNAMIC OBFUSCATION
The map must expire before the chain completes.
Static defense-in-depth collapses against autonomous chain reasoning. Infrastructure must mutate — hyper-segmentation, ephemeral compute, continuous key rotation.
SYNTHETIC IDENTITY VALIDATION
Authorization is no longer the question.
The question is whether the entity in the session is human — or a synthetic agent three steps into a chained exploit. Behavioral telemetry recalibrated for AI cadence.
We thought we were building a conscience. It turned out we were also building a microscope.
Teaching a machine to find the exact edge where a rule breaks is teaching it, in some very deep sense, to find the exact edge where anything breaks. The immune system was a side effect.
ACRA INSIGHT // CONTEXT JAMMING
Field reporting draws on Anthropic's Project Glasswing disclosures, the Claude Mythos Preview System Card, reporting from VentureBeat, the Guardian, the Washington Post, Lawfare, and SecurityBoulevard, and evaluations published by the UK AI Security Institute.
- · Anthropic · Mythos Preview system card
- · UK AI Security Institute · expert CTF protocol
- · SWE-bench Pro / Verified · benchmark set
- · CyberGym · vulnerability reproduction suite
- · 10 U.S.C. § 3252 · supply-chain designation filings