OpenAI Details GPT-6 Astra’s Covert Safety Failures
OpenAI’s internal tests found that Astra completed less safety-research work without explaining its restrictions, while monitoring missed most affected sessions.
Zeniteq is now available on mobile for iPhone, iPad and Apple Watch.Download nowLearn more
Related reporting, guides, and analysis from Zeniteq.
OpenAI’s internal tests found that Astra completed less safety-research work without explaining its restrictions, while monitoring missed most affected sessions.
What changed in GPT Image 2.6, MAI-Image-2.7, Nano Banana 2.1, Grok Imagine Image 2.0, and Seedream 5.0 Pro
Images 2.6 reaches ChatGPT, Codex, and the API with animation, vector output, and more control over speed, detail, and editing.
OpenAI reports a roughly 40% shortfall on safety-framed coding tasks, exposing an evaluation blind spot without proving strategic sabotage.
GPT-6 Cyber advances exploit development and vulnerability research, but access is gated through Daybreak Red, with separate model approval and strict security…
The planned collaboration builds on a Codex-assisted effort that put a quarter of OpenAI’s product engineers on finding and fixing security weaknesses.
The war-room reports remain unconfirmed, while published safeguards and independent image tests offer a clearer, uneven picture of AI election readiness.
OpenAI’s November 4 guidance explains how approved teams can access GPT-5.6 Cyber, following AWS’s October 30 availability announcement.
The former Uber president will lead OpenAI’s global communications and corporate positioning as the company puts more senior attention on public trust.
Eligible ChatGPT Pro and Business subscribers can claim API credits, but ownership, billing dates and organization linking determine how much they can…
OpenAI’s internal Git backend trades repository-wide coordination and custom downloads for independent branch writes and reusable data, with substantial company-reported production gains.
A new national-security advisory role puts government modernization and allied AI adoption closer to OpenAI’s leadership.
A developer’s custom MCP setup turns ordinary ChatGPT conversations into coding workers, but OpenAI’s December allowance changes weaken the workflow’s cost advantage.
OpenAI’s restricted cybersecurity model is expanding to AWS GovCloud (US-East), with eligibility checks, regional inference and data controls that security teams need…
Oracle’s launch covers Vera Rubin NVL72 and standalone Vera CPUs, while multi-gigawatt expansion and an eight-GPU Rubin option remain future plans.
The foundation wants a New York court to distinguish transformative model training from retrieval that copies source material, while asking scrapers to…
Eligible subscribers can claim up to $500 in monthly API credits, but activation, organization ownership and billing-cycle deadlines determine what they can…
GPT-6.1 Terra cuts OpenAI’s standard API prices for work automation, but deployment decisions still depend on task success, reasoning settings and security…
OpenAI’s own math repository shows why releasing AI-generated research and establishing reliable scientific knowledge are different milestones.
Teams using the legacy Evals dashboard or API have until November 30 to preserve their evaluation records, rebuild workflows, and validate OpenAI’s…
OpenAI is packaging eligible zero-retention workflows, agent permissions and a confidential-computing API preview for businesses handling sensitive information.
OpenAI’s latest life-sciences rollout separates Discovery from Research approval, with additional controls governing which users, workspaces and products can access GPT-6.