· 2 min read
The Gradient — 2026-09-19
- Industry News
- Company Blog
Security researchers demonstrated that Anthropic’s Claude could be weaponized to infiltrate OpenAI’s internal infrastructure, exposing critical gaps in AI safety.
Researchers used Anthropic’s Claude to hack into OpenAI
What happened: Security researchers leveraged Anthropic’s Claude to identify and exploit several undocumented vulnerabilities in OpenAI’s internal tools. By prompting Claude to generate malicious code, they hijacked employee accounts and extracted a private code repository before responsibly disclosing the findings.
Why it matters: The incident underscores how powerful language models can be turned into attack vectors, raising concerns about AI supply‑chain security, cross‑company trust, and the need for stricter model‑usage controls.
Key stats:
- 3 employee accounts compromised
- 1 private code repository accessed (≈ 2 GB)
- Flaws reported within 48 hours of discovery
Source: https://techcrunch.com/2026/09/18/researchers-used-anthropics-claude-to-hack-into-openai/
Hex turns complex analysis into visual reports with GPT‑6 Astra
What happened: Hex integrated the GPT‑6 Astra model into its Hex platform, enabling data agents to turn textual answers into interactive visual reports that employees can easily share.
Why it matters: By automating the creation of polished visualizations, teams spend less time formatting data and more time deriving insights, boosting productivity and cross‑team collaboration.
Key stats: The announcement did not disclose specific performance metrics or adoption numbers.
Introducing the Australian Youth Safety Blueprint
What happened: OpenAI released the Australian Youth Safety Blueprint, a six‑pillar framework that outlines best practices, safeguards, and empowerment tools for AI products aimed at users under 18 in Australia.
Why it matters: Youth are among the most vulnerable AI users, and the Blueprint provides a concrete, region‑specific model for responsible deployment, helping regulators, developers, and educators align on safety standards.
Key stats:
- Six pillars cover transparency, consent, education, parental controls, bias mitigation, and incident response.
- Targets AI interactions for roughly 5 million Australian minors.
- Sets measurable goals for reducing harmful content exposure by 30% within the first year.
Source: https://openai.com/index/australian-youth-safety-blueprint
How Fyxer built an AI executive assistant people trust
What happened: Fyxer launched an AI‑driven executive assistant that organizes inboxes, drafts replies, and learns each user’s voice through continuous fine‑tuning and memory features.
Why it matters: By delivering trustworthy, voice‑consistent communications, the tool reduces executive overload, speeds up decision‑making, and sets a new bar for AI‑augmented productivity.
Key stats:
- Over 10,000 early‑adopter users
- Average email‑drafting time cut by 3×
- 95% user satisfaction rating on trust and accuracy
Source: https://openai.com/index/fyxer
Perplexity trusts GPT-6 Astra with end-to-end systems
What happened: Perplexity integrated the new GPT‑6 Astra model, dubbed “Astra,” to autonomously draft internal communications, modify software, and oversee live production systems. The company reports that it now checks in on these processes much less frequently than with previous models.
Why it matters: Automating these critical tasks cuts operational overhead, speeds up iteration cycles, and reduces human error. The shift signals confidence in LLMs handling complex, end‑to‑end workflows in real‑time environments.
Key stats: • Check‑in frequency reduced dramatically compared to earlier models (exact numbers not disclosed). • Astra is the first GPT‑6 variant deployed at scale for full‑stack operations.
Source: https://openai.com/index/perplexity-improving-accuracy-with-astra