“If man chooses oblivion, he can go right on leaving his fate to his political leaders. If he chooses Utopia, he must initiate an enormous education program - immediately, if not sooner.”
-R B Fuiler

  • 1 Post
  • 12 Comments
Joined 10 days ago
Cake day: August 25th, 2025
  • nymnympseudonym@piefed.socialtoblueteamsec@infosec.pub•AI-powered ransomware named PromptLock - uses the gpt-oss:20b model from OpenAI locally via the Ollama API to generate malicious Lua scripts on the flyEnglish
    11·
    9 days ago

    Perplexity prompt: “quotes from openai about safety testing for gpt-oss”

    Reply below:

    Safety Testing Measures

    • OpenAI’s safety process for open-weight release included technical audits (robustness, bias detection), structured adversarial testing, and evaluation of prompt-injection and jailbreak vulnerabilities.
    • The organization also employed both automated and human review to test for outputs that could cause harm, leveraging cross-disciplinary teams including outside researchers, ethicists, and cybersecurity experts.
    • Feedback from these rounds led to incremental model improvements before the open weights went public.

    Transparency and External Collaboration

    • OpenAI has collaborated with third-party security and ethics researchers to validate its safety protocols and stress-test new models prior to release.
    • The company acknowledges that “Releasing open weights is a significant responsibility due to risks of misuse. We want to be transparent about our process and invite the AI community to help report and address issues that may arise post-release.”