Last week, Anthropic released Mythos, stating it would only be available to 40 organizations, helping some teams uncover a 27-year-old vulnerability. This week, OpenAI fired back directly: GPT-5.4-Cyber is here, targeting thousands of security researchers and hundreds of teams, dwarfing Mythos's elite club approach in scale.
The two companies are competing in the same arena within the same time window, using two completely different strategies.
What is GPT-5.4-Cyber?
This is not a brand-new model, but a cybersecurity-specific variant of GPT-5.4. The core difference is that restrictions for defensive security scenarios have been selectively relaxed.
The most critical new capability: binary reverse engineering.
Typical security analysis often requires access to source code or manual disassembly. GPT-5.4-Cyber can directly accept compiled binary files to help find vulnerabilities and identify malicious logic. In reality, only a minority of malware comes with source code available—much APT attack analysis, ransomware reverse engineering, and supply chain security auditing deals with compiled black boxes. This feature directly addresses the most painful scenarios for security engineers.
From 27% to 76%: What These Numbers Mean
OpenAI used the CTF (Capture The Flag) benchmark to measure progress in security capabilities:
| Time | Model | CTF Benchmark Score |
|---|---|---|
| August 2025 | GPT-5 | 27% |
| November 2025 | GPT-5.1-Codex-Max | 76% |
Jumping from 27% to 76% in three months means AI's practical problem-solving ability in cybersecurity has nearly tripled. A year ago, AI could only handle basic security Q&A; now it can solve most medium-difficulty CTF challenges.
Meanwhile, the code security analysis tool Codex Security, based on GPT-5.4-Cyber, has helped fix over 3,000 high-severity or critical vulnerabilities since its private beta.
The TAC Program: Tiered Access, Not Open to All
Access to GPT-5.4-Cyber is through the Trusted Access for Cyber (TAC) program, which launched in February alongside a $10 million cybersecurity grant, initially running on GPT-5.3-Codex.
This upgrade introduces a multi-tier identity verification system:
- Basic tier: Standard verification, access to general cybersecurity features
- Advanced tier: Stricter identity verification, unlocks stronger capabilities
- Top tier: Highest verification level, required for GPT-5.4-Cyber access
Individual registration is available at chatgpt.com/cyber, while enterprise teams go through OpenAI's sales channel. Existing TAC members can apply individually for upgrades to higher tiers.
Anthropic vs OpenAI: A Direct Clash of Strategies
The timing is too coincidental to be accidental.
Last week, Anthropic released Mythos, emphasizing an elite approach, a small number of partners, and results-driven validation. OpenAI's response is a broad rollout, involving thousands of participants, winning through coverage.
| Anthropic Mythos | OpenAI GPT-5.4-Cyber | |
|---|---|---|
| Release Date | First week of April 2026 | April 14, 2026 |
| Target User Scale | ~40 organizations | Thousands of individuals + hundreds of teams |
| Core Strategy | Elite, prove capability with results | Broad rollout, create tangible defensive value |
| Signature Capability | Uncovered a 27-year-old vulnerability | Binary reverse engineering, 76% CTF benchmark |
Which is more effective? It's too early to say. But both companies racing for the AI cybersecurity track within the same time window already shows this direction is being treated as a true strategic battleground.
Another Notable Boundary
GPT-5.4-Cyber's relaxed restrictions have limits, applying only to defensive security scenarios. OpenAI requires stricter deployment controls, and an agreement must be signed before use.
Analyzing the C2 logic of malware: allowed. Using it to write new attack tools: definitely not.
How these boundaries will be implemented in practice still awaits community feedback from actual use.
Sources: OpenAI launches GPT-5.4-Cyber model for vetted security professionals (SiliconANGLE); CocoLoop, OpenAI Releases Cyber Model to Limited Group in Race With Mythos (Bloomberg)