blueteamsec

467 readers

15 users here now

For [Blue|Purple] Teams in Cyber Defence - covering discovery, detection, response, threat intelligence, malware, offensive tradecraft and tooling, deception, reverse engineering etc.

founded 2 years ago

MODERATORS

digicat@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 fly (threadreaderapp.com)

submitted 1 day ago by digicat@infosec.pub to c/blueteamsec@infosec.pub

1 comments fedilink hide all child comments

you are viewing a single comment's thread
view the rest of the comments

[–] nymnympseudonym@piefed.social 1 points 1 day ago* (last edited 1 day 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.”