Breaking

OpenAI Says GPT-6 Astra Found Zero-Days as Monitorability Fell

September 8, 2026 at 13:03 EDT

Security
Foundation Models
AI Agents

OpenAI says GPT-6 Astra discovered previously unknown vulnerabilities and built working exploit chains in controlled lab evaluations, while its safety disclosures acknowledge a significant decline in chain-of-thought monitorability compared with GPT-5.6 Sol. The combination gives the model a stronger defensive-security use case, but raises new questions about whether oversight systems can reliably detect harmful behavior.

September 3, 2026 · OpenAI

GPT-6 Astra finds zero-days

OpenAI says Astra discovered previously unknown vulnerabilities in controlled lab tests. Its safety disclosures also say the model is harder to monitor through written reasoning than GPT-5.6 Sol.

One 22-target lab evaluation

Successful targets

GPT-6 Astra
10/22
45.5%

GPT-5.6 Sol
1/22
4.5%

The evaluation contained 22 isolated targets; all but one vulnerability was a zero-day.

Cyber threshold

Critical

Findings disclosed

ExploitBench

100%

41 V8 vulnerabilities

OS evaluation

12 h

Local privilege escalation

How oversight works

Reasoning and actions

→

Async monitor

→

Pause or end

The monitor can inspect the full trajectory, but OpenAI says it may still miss misaligned behavior before intervention.

Capability

Finds unknown flaws in long-horizon tests

Builds exploit chains with tools

Monitoring risk

Written reasoning can be less revealing

CoT-only monitors can miss evasive behavior

Bottom line

Safety is layered, not solved

Astra is more aligned overall, but OpenAI warns that monitoring can miss harmful behavior. The deployment combines reasoning review, action monitoring and access controls to manage that gap.

Continue reading

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