[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← AI · markdown

confirmed (vendor research)

White Circle RISE: strategy-level red-teaming bypasses modern image model safety; newer GPT-Image/Nano Banana worse

White Circle published RISE (Red-teaming via Iterative Strategy Evolution), an open pipeline that evolves reusable natural-language attack strategies with an abliterated Qwen3-32B attacker and a human-calibrated Gemini judge, rather than paraphrasing seed prompts. On human-verified nudity attacks they report up to 13.3% ASR on DALL·E 3 after exploitation; post-publication runs on GPT-Image-2.5-Flare and Nano Banana 2.1 roughly doubled ASR versus prior versions (7.8% and 20.7% nudity). Prior methods that claimed ~30%+ success fell near 0% under the same calibrated judge. Code is at github.com/whitecircle/rise; strategies, attack prompts, and unsafe images are withheld. Paper: arXiv:2609.34920. Findings were shared with affected vendors before publication.

Sources: White Circle RISE · arXiv:2609.34920 · GitHub whitecircle/rise