RISE Finds New Image Models Easier to Jailbreak

// 1h agoBENCHMARK RESULT

RISE Finds New Image Models Easier to Jailbreak

White Circle says GPT-Image-2.5 Flare and Nano Banana 2.1 appear more breakable than their predecessors under the same red-team process. However, the published RISE preprint reports human-verified rates only for earlier models, reaching up to 13%.

// ANALYSIS

The real story is evaluation quality: newer image models may be trading safety margin for capability, but the latest comparison still needs independently published numbers.

–RISE evolves reusable attack strategies instead of endlessly rewriting individual prompts.
–Its calibrated judging reportedly drives earlier methods’ claimed success rates near zero.
–The public evidence covers DALL·E 3, Nano Banana 2, and GPT-Image-2—not yet the newer 2.5 and 2.1 models.
–Developers shipping image generation should test safety regressions after every model upgrade, not assume newer means safer.
–The gap between vendor safety cards and independent red-team results makes standardized external evaluations increasingly important.

// TAGS

risebenchmarkevaluationsecuritysafetyimage-genresearch

DISCOVERED

1h ago

2026-10-14

PUBLISHED

1h ago

2026-10-14

RELEVANCE

9/ 10

AUTHOR

TTrimoreau

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

VIDEO42m ago

ChatGPT Work Turns Catering Goals Into Deliverables

OpenAI demonstrates ChatGPT Work helping a Toronto restaurant research catering prospects, build a website and tracker, and produce a branded menu. The workflow shows an agent turning a broad business objective into reviewable, ready-to-use assets.

LAUNCH54m ago

Composio Instant Gives Agents Tool Wallets

Composio Instant lets agents access 100+ paid tools across search, scraping, enrichment, voice, transcription, and media without separate provider accounts, API keys, or subscriptions. Calls draw from a shared Composio balance across Claude, ChatGPT, Codex, Cursor, and other MCP clients. [Composio](https://composio.dev/blog/composio-instant)

NEWS56m ago

Modelers.cn batches a dozen Ascend quants

Modelers.cn is surfacing a burst of Ascend-targeted model repositories, most quantized variants of models trained elsewhere. The trend shows open-weight distribution shifting toward hardware-specific packaging and deployment optimization.