AI Cybersecurity Just Got Autonomous: Mythos, Glasswing, GPT-5.4-Cyber

Ten days, three shipped models, one $190 million check, and a stock chart that looks like a trapdoor. April 2026 is when AI in cybersecurity stopped being a conference keynote slide and started being the product cycle. If you ship software, defend a network, or run anything that touches the public internet, the threat model this month is not the one you planned for last quarter.

Here is what actually happened, in order, with links. No "game changer" language, no breathless futurism, just what shipped and what it means for the people whose job is to ship code or keep it from being broken.

Project Glasswing and Claude Mythos Preview

On April 7, 2026, Anthropic published the Mythos Preview announcement and launched Project Glasswing. Mythos Preview is a new general-purpose Claude model. Glasswing is the private program Anthropic built around it because the company decided not to ship the model publicly.

The model found zero-day vulnerabilities in every major operating system and every major web browser. Thousands of them, including high and critical severity bugs, some of them 27 years old. Anthropic reports that in comparable Firefox testing, Opus 4.6 had a near-zero autonomous exploit development rate. Mythos Preview hit 181 successful exploits in the same harness. Running the model at roughly $10,000 of API credits against FFmpeg and $20,000 against OpenBSD produced hundreds to thousands of runs, respectively, each capable of chaining multiple vulnerabilities, reverse-engineering closed-source binaries, and writing working exploits without a human in the loop.

The Project Glasswing partner list reads like a Fortune 500 who-has-root list: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic said around 40 additional critical-infrastructure organizations are also inside the program. The company is committing up to $100 million in Mythos Preview usage credits to Glasswing partners and another $4 million in direct donations to open-source security orgs.

The position is explicit: Mythos Preview will not be generally available. Anthropic believes it is too dangerous to ship. That judgment alone is the story. A frontier lab shipped a model good enough at offensive security that it decided to give it only to the defenders of the most critical software on earth. US Treasury Secretary and Federal Reserve Chair reportedly warned bank CEOs about the cyber risks the same week.

Claude Opus 4.7 and the Cyber Verification Program

On April 16, Anthropic shipped Claude Opus 4.7 for general use, and with it a new safeguard layer. Opus 4.7 ships with real-time cyber safeguards that automatically detect and block two kinds of requests: prohibited uses like ransomware code or mass data exfiltration, and high-risk dual-use activities like vulnerability testing and exploit development that also have legitimate defensive applications.

If your day job is pentesting, red-teaming, or vulnerability research, your day job just got blocked by the model you were paying for. Anthropic's answer is the Cyber Verification Program. It is a free, application-based allowlist. You submit the Cyber Use Case Form, document your intent, and Anthropic reviews within two business days. Approved users get the safeguards relaxed for dual-use work. Unapproved users keep hitting the block.

Access routes differ by platform. Direct Claude users (Claude.ai, Claude Code, the API) apply directly. Microsoft Foundry users select Azure and provide tenant and subscription IDs. Amazon Bedrock and Google Vertex AI do not currently support the program. Zero Data Retention customers have to go through their Anthropic sales rep. Per The Register, ID verification for the program runs through Persona.

The short version: Anthropic just built a registration gate around cyber capabilities. Legitimate security pros can still get through, but the vibes-coded weekend-hacker path is closed.

OpenAI ships GPT-5.4-Cyber

Two days before Opus 4.7, on April 14, OpenAI launched GPT-5.4-Cyber, a fine-tuned variant of GPT-5.4 with deliberately lowered refusal boundaries for defensive cybersecurity work. The model adds binary reverse engineering for malware analysis and vulnerability discovery, plus broader capability to reason about offensive techniques that the stock GPT-5.4 refuses outright.

Like Anthropic's model, GPT-5.4-Cyber is not generally available. OpenAI scaled up its Trusted Access for Cyber (TAC) program to thousands of verified individual defenders and hundreds of teams, and gates the cyber model behind the highest TAC tiers. Two competing frontier labs, shipping two cyber-specialized models, inside one week, both gated behind verification programs. That is not a coincidence; that is an industry admitting that the unmodified models can cause too much trouble to leave open.

The defender problem

On March 10, Kevin Mandia, the founder of Mandiant (sold to Google for $5.4 billion in 2022), came out of retirement with Armadin, an AI-native cybersecurity startup. Armadin closed $189.9 million in combined seed and Series A, led by Accel, with GV, Kleiner Perkins, Menlo Ventures, 8VC, Ballistic Ventures, and In-Q-Tel. The co-founders are Travis Lanham (ex-Google Cloud Security), Evan Pena (ex-Mandiant), and David Slater (ex-Google SecOps).

The pitch is one sentence: autonomous cybersecurity agents that learn and respond to threats without a human in the middle. Mandia's framing of the threat model: "When you have AI on offense, what you are going to get is a technology that can think, can learn, can adapt." Attackers, he warned, will complete attacks in minutes that used to take days.

Anthropic's own threat reporting backs this up. The GTG-1002 campaign used autonomous agents to automate 90 percent of the intrusion lifecycle, collapsing the window between initial access and impact from hours to, in some measured cases, 22 seconds. A human SOC analyst reading a Slack alert cannot compete with that. An agent running a playbook can, which is why the defensive side of the market just raised a record seed round to build those agents.

The CrowdStrike hedge

Not every incumbent is buying it. CrowdStrike CEO George Kurtz argues that while AI agents can reason about code, they cannot manage the operational chaos of a live enterprise environment. Markets reportedly took that as a defensive statement more than a confident one; CrowdStrike's stock took a hit in February as investors reassessed the moat around traditional endpoint-detection-and-response platforms.

Kurtz is not wrong that agents can stumble in messy real-world environments. He is also in the position of a CEO whose product line was designed for an earlier threat model. The honest read is that some of the incumbents will pivot successfully and some will not, and the ones that move first on agentic defense have the better odds.

What this means for open-source

Nothing about the frontier closed-access play changes the open-source picture yet. SmolLM3, Qwen3.5, Gemma 4, and the rest of the fully open stack still run with standard safeguards. Community-maintained tooling like Anthropic-Cybersecurity-Skills, a 754-skill library mapped to MITRE ATT&CK, NIST CSF 2.0, D3FEND, and NIST AI RMF, is pushing structured cyber workflows into any agent framework that understands the agentskills.io standard.

But the gap is widening. Mythos Preview and GPT-5.4-Cyber are qualitatively better at offense than anything you can download. If the pattern holds, future frontier cyber models will stay locked behind verification programs, which means open-source defense will have to ride on open-source models that lag the frontier by six to twelve months. Defenders running fully open stacks will need to invest harder in tooling, telemetry, and playbooks, because the raw model horsepower gap is not going to close itself.

What to do this week

Four concrete moves if you work in or around software security:

• If you do legitimate cyber work with Claude, apply to the Cyber Verification Program now. Two business days is the stated turnaround. Opus 4.7 will keep blocking your prompts until you are through.
• If you ship critical software and are not already in Project Glasswing, watch for Anthropic's outreach. The company says there are about 40 organizations beyond the named partners already inside; the list is not closed.
• If you run a SOC, start pressure-testing your response playbooks against a hypothetical 22-second agent-driven intrusion. Most existing playbooks assume human-in-the-loop containment at minutes-to-hours timescales. That assumption is obsolete for the new threat profile.
• If you build with open-source models, assume you are six months behind the frontier on offensive capability and plan defense accordingly. Invest in behavioral detection, hardening, and response automation that does not depend on a frontier model behind a verification gate.

The verdict

The thing that makes this week different is not any one model. It is the combined signal. Two frontier labs shipped cyber-specialized models in seven days. Both gated them behind verification programs. The founder of Mandiant came back with nine figures to build autonomous defense agents. The federal government started calling bank CEOs. Anthropic found thousands of zero-days in software you use every day and decided it was safer not to ship the model that found them.

The cybersecurity industry spent the last decade arguing about whether AI would matter for defense. That argument is over. Whether your stack is ready for the next six months of this is a different question, and one worth answering before something tests it for you.

Sources

• Anthropic: Claude Mythos Preview announcement (April 7, 2026)
• Anthropic: Project Glasswing
• Anthropic: Zero-day findings from Mythos Preview
• Anthropic: Real-time cyber safeguards on Claude
• Help Net Security: Claude Opus 4.7 released with cyber safeguards
• OpenAI: Trusted access for the next era of cyber defense
• 9to5Mac: OpenAI unveils GPT-5.4-Cyber
• TechCrunch: Kevin Mandia raises $190M for Armadin
• IBM: Mythos and Glasswing raise the stakes
• PostQuantum: Mythos Preview and the end of a twenty-year equilibrium