Breaking

OpenAI Limits Astra Cyber Access as Anthropic Updates Claude Models Ahead of IPOs

September 2, 2026 at 05:16 EDT

Foundation Models
AI Agents
Security

OpenAI is preparing a broad release of Astra while reserving its most advanced cybersecurity features for a small group of testers. Anthropic, meanwhile, has released Claude Fable 5.1 and Claude Mythos 5.1 with lower operating costs and fewer safeguard interventions, highlighting two different approaches to managing frontier-model risk as both companies pursue public-market financing.

SEPTEMBER 2, 2026 · OPENAI + ANTHROPIC

Safety and speed split the frontier

OpenAI is staging Astra's highest-risk cyber capabilities, while Anthropic is broadening Fable 5.1 with fewer safeguard interventions.

Four signals behind the shift

zero-days chained by Astra

85%

fewer biology interventions

~60%

fewer cyber interventions

75%

lower cache-read cost

Safeguards are becoming more selective

Reported reductions in interventions, shown as proportional vertical blocks.

85%

Medical / biology

~60%

Cybersecurity

The largest reduction is in medical and biology interventions.

THE RELEASE TRADE-OFF

More capability is arriving through more controlled access.

OpenAI

Advanced Astra cyber access starts with testers.

Anthropic

Mythos 5.1 remains limited to vetted organizations.

IPO clock

Both labs must show growth while addressing risk.

Useful progress, unresolved trade-offs

Potential upside

Lower friction

Fewer interventions can help legitimate biology and security work continue.

Remaining constraint

Residual risk

Astra safeguards can still flag legitimate tasks and stop API work.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.