How AI Safety Model Hype Turns Anxiety Into Business

In April 2026, the cybersecurity community was flooded with two headlines—Anthropic claimed Claude discovered “thousands of zero‑day vulnerabilities” but was too dangerous to release; OpenAI responded with a “defense‑focused” GPT‑5.4‑Cyber. The article argues these claims are largely hype.

1. “Too dangerous to release” – a familiar tactic

On April 7, Anthropic published a technical report for Claude Mythos Preview, stating the model could discover “thousands of zero‑day vulnerabilities” and perform a simulated penetration test within hours. The company then declared the technology “too dangerous to publish”. Only twelve partners and over forty critical‑infrastructure organizations can access it, leaving the public excluded.

The move mirrors earlier strategies: Anthropic’s CEO Dario Amodei previously worked at OpenAI, which in 2019 labeled GPT‑2 “too risky to release”. Sora, a video‑generation model, was similarly held back and later quietly discontinued.

Anthropic now frames the narrative as “general AI safety” shifting to “cybersecurity”, but the core script remains the same: the most powerful weapon is to withhold the product.

2. Data tells the truth: not that spectacular

The UK Government AI Safety Institute (AISI) independently evaluated Claude Mythos. It performed well on basic tasks (completion rate >85%) but was comparable to other frontier models—GPT‑5.4, Opus 4.6, and Codex 5.3—all on a similar performance level.

The most notable result came from the multi‑stage intrusion simulation (TLO test). Mythos was the first model to complete all 32 steps of the simulated attack, but this success occurred only in a “few attempts”. On average, the model completed 22 steps, whereas the previous generation averaged 16 steps.

In a more complex industrial‑control‑system interference simulation, Mythos performed poorly, indicating limited capability in real‑world, intricate network‑security scenarios.

3. OpenAI’s “defensive counter‑attack”: you attack, we defend

One week after Anthropic’s announcement, OpenAI launched the defense‑oriented GPT‑5.4‑Cyber. The timing suggests deliberate “benchmark marketing”: each company tells its own story to capture market attention and budget.

Tech media bluntly observed that large firms can release new models and still easily steer public sentiment, using anxiety‑driven marketing to “cut investors like leeks”.

When the industry debates “Anthropic vs OpenAI”, the actual technical gap becomes secondary; the marketing narrative keeps audiences busy choosing sides instead of questioning the underlying technology.

4. “Security alliance” as a policy passport

Alongside Claude Mythos, Anthropic announced the “Glasswing” initiative, gathering Microsoft, Apple, AWS, Nvidia, and CrowdStrike into a defensive alliance. Anthropic contributed $1 billion and raised $4 million for open‑source security projects.

This move serves a compliance purpose: by demonstrating industry self‑regulation before legislation, Anthropic seeks a regulatory “passport” while excluding competitors lacking alliance membership.

Thus, “security” shifts from a technical goal to a competitive barrier and policy lever.

5. Old script, new stage: recurring AI‑security hype

Exaggerated marketing in AI security is not new. Months before Claude Mythos, the UK research group Giskard reported systematic over‑statement in AI‑security assessments, noting that hundreds of tests suffer from serious flaws that weaken derived conclusions.

Security experts note that large‑language‑model safety tests often rely on limited, known attack patterns that do not reflect the diverse real‑world adversarial environment.

Forrester’s security analysis bluntly states that the cybersecurity industry over‑emphasizes “new attacks” and “zero‑day bugs”, while neglecting that such attacks are rarely essential in practice.

Even if AI could autonomously discover zero‑days, attackers still favor cheaper methods like phishing and social engineering, which remain effective today.

6. Technology is real, but don’t over‑believe

The article does not dismiss AI progress in cybersecurity. Mythos’s ability to complete the full 32‑step TLO simulation shows improved agent‑style planning. Google’s Big Sleep project previously demonstrated AI‑driven discovery of a zero‑day in SQLite, confirming that AI‑assisted vulnerability hunting is feasible.

However, the line between genuine technical advancement and marketing narrative must be clearly drawn. When a claim such as “found thousands of zero‑days covering all major systems” is packaged from a single conditional discovery, or when an average of 22 steps is spun into “too powerful to release”, the result is marketing, not technical discourse.

Conclusion: Clarity is the rarest security skill

From GPT‑2 to Sora, from GPT‑5.4‑Cyber to Claude Mythos Preview, the AI‑security marketing playbook repeats:

Manufacture scarcity

Declare the technology “too dangerous”

Form a “security alliance” with core partners

Stoke market anxiety to drive stock volatility

Compared with today’s panic narrative of “digital security on the brink of collapse”, history shows that reality usually lies between exaggerated hype and actual impact.

Faced with this carefully choreographed “AI security arms race”, CISOs, security practitioners, and investors should return to basic questions:

What can the technology truly do?

What are its limitations?

How much of the marketing is independently verifiable fact versus crafted story?

In an era where anxiety becomes a commodity, staying sober‑minded is the scarcest security capability.