OpenAI's Cyber Gatekeeping Flip-Flop: Altman Slams Anthropic, Then Locks Down Rival Tool

Sam Altman didn't hold back. He called Anthropic's restrictions on its Mythos cybersecurity model "fear-based marketing." Now OpenAI is doing the exact same thing with its own powerhouse, GPT-5.5 Cyber. Irony bites hard in AI's cutthroat arena.

The TechCrunch report dropped April 30, 2026, laying it bare: After Altman trashed Anthropic for gatekeeping Mythos to select users, he posted on X that OpenAI would roll out GPT-5.5 Cyber "to critical cyber defenders" in days. TechCrunch. Applicants must submit credentials and use plans via OpenAI's form. Cyber handles penetration testing. Spots vulnerabilities -- and exploits them. Reverse-engineers malware. A defender's dream. Or an attacker's blueprint.

This isn't OpenAI's first dance with limits. Back in February, they launched Trusted Access for Cyber (TAC). Vets pros for riskier models. By mid-April, they scaled it, introducing GPT-5.4-Cyber to hundreds, then thousands of verified defenders. OpenAI blog. The New York Times noted the parallel: A week after Anthropic's Mythos lockdown, OpenAI followed suit, sharing GPT-5.4-Cyber only with partners. New York Times. Tools like these find software holes. Defend networks. But flip the script -- they breach them too.

Anthropic kicked it off April 7. Unveiled Claude Mythos Preview. Too potent for public eyes. Capable of unearthing thousands of zero-days autonomously. They formed Project Glasswing: Invite-only for 40 orgs like JPMorgan, AWS, CrowdStrike. The Economist. Exclusivity sparked envy. Asian bank boards demanded in. But leaks happened. Unauthorized users breached Mythos via blunders and data spills, per Bloomberg and Tom's Hardware. Tom's Hardware.

Why the Lockdowns? Dual-Use Dilemma at Center Stage.

Cybersecurity AI walks a razor's edge. Mythos and Cyber excel at offense and defense. Anthropic warned Mythos could supercharge attacks if loose. OpenAI echoes that, consulting U.S. government for broader trusted access. The Verge reported Altman's X post Thursday: Work with ecosystem and feds on Cyber rollout. The Verge. SecurityWeek detailed OpenAI widening TAC post-Mythos, to thousands. But speed matters. Time-to-exploit shrinks. Remediation lags.

Governments scramble. Central banks, intel agencies buzzed after Mythos news. The Guardian highlighted fears: Mythos IDs unknown IT flaws ripe for hackers. The Guardian. OpenAI briefed U.S. feds, states, Five Eyes on GPT-5.4-Cyber. Axios scooped demos to 50 gov cyber pros. Tiered access: Vetting matches commercial, plus stricter safeguards.

Industry watches. Wired covered OpenAI's strategy shift: Safeguards cut cyber risk for now; Cyber fills the gap. Wired. Critics smell commerce. The Economist: Locking top models hands power to makers. Envy drives bids. But hypocrisy stings. OpenAI, once open-ish, now hoards like rivals.

Recent X chatter amplifies. TechCrunch's post hit 9K views, sparking shares. Users note Altman's shade-throwing backfire. One called it "classic OpenAI hypocrisy." Kenshi posted: OpenAI locks models, labels it safety -- echoing Economist on elite access.

Risk, Rivalry, and the Road Ahead.

OpenAI plans expansion. Gov talks. Credential checks. But breaches like Mythos expose cracks. Unauthorized Discord claims: Mythos accessed despite walls. Firebox bugs fixed by it anyway.

Competition fuels this. Anthropic's Glasswing coalition includes OpenAI foes. OpenAI counters with TAC tiers. Market? $35B AI cyber spend incoming. Medium pegged it: New "restricted-access defensive AI" category born.

Altman pivoted fast. From dissing to mirroring. Safety trumps swagger when exploits loom. Defenders gain edge -- if access holds. Attackers? They'll pry anyway. And the arms race? Just heating up.