Open-source RISE red-teaming tool claims to break image model safeguards at OpenAI and DeepMind
dair_ai · x · 2026-10-15
WhiteCircle has open-sourced RISE, a red-teaming algorithm that reportedly breaks the safeguards of the latest image models from OpenAI, DeepMind, and others.
Key points:
- RISE learns better ways to write attack prompts, with no additional training required
- It runs on readily available models: open-source Qwen generates attacks while Gemini Flash acts as the judge
- Fully open source, so defenders can run the same tests against their own image models
The framing: what AI Safety needs is better tooling and engineering, and RISE is a working demonstration of that thesis.
Related event: Open-Source Red-Teaming Tool RISE Breaks Image Model Guardrails(3 posts)→
More from Safety
- Leaked Claude and OpenAI keys for 4 days: only keys inside running pods were the ones leaking — Ubunta · 2026-10-15
- Banking regulators ramp up AI scrutiny as rogue agents threaten financial stability — nordicinst · 2026-10-15
- Anthropic acqui-hires Opaque Systems team, licenses its confidential computing tech for AI — rohanpaul_ai · 2026-10-15
- Safety researcher quips: one spy found in your AI company means there are more — dhadfieldmenell · 2026-10-15
- promptspeak-mcp-server ships 45 MCP tools for pre-execution agent governance: hold queues, audit trails, risk scoring — modelcontextprotocol · 2026-10-15
- EU-China talks aim to save ASML supplier Sindlhauser from export-control squeeze — teortaxesTex · 2026-10-15