Monitor this situation.
Unsubscribe anytime.
[SITUATION] · [QUIET] · [TECHNOLOGY]
2 clusters · 4 sources · 15 days · First seen · Last updated
OpenAI and Google AI model safety vulnerabilities
Overview
OpenAI has disclosed a “self-replicating prompt injection” vulnerability discovered during adversarial training of GPT-5.6 models. This vulnerability allows malicious instructions to mimic computer worm behavior by inducing models to repeat injections in subsequent outputs, such as in email replies or Excel files. OpenAI reported these instances occurred only in controlled environments with no real-world incidents.
Further research indicates that advanced image models from OpenAI and Google remain susceptible to adversarial prompts designed to bypass safety guardrails. Testing showed that OpenAI’s GPT-Image-2.5 Flare and Google’s Nano Banana 2.1 were vulnerable to prompts seeking explicit content. Studies also noted that “adversarial poetry” and minor prompt adjustments could elicit forbidden imagery. These vulnerabilities are linked to broader social impacts, including reports of image-based sexual abuse and exposure to AI-generated sexual material among teenagers.
Entities
Timeline
-
[TECHNOLOGY] 2 sourcesOpenAI and Google AI models vulnerable to explicit content prompts
Researchers have found that OpenAI and Google AI models can be bypassed using adversarial prompts to generate explicit content, posing risks to users and teenagers.
-
[TECHNOLOGY] 2 sourcesOpenAI discloses self-replicating prompt injection vulnerability
OpenAI has identified a “self-replicating prompt injection” vulnerability in GPT models that allows malicious instructions to spread like a worm through AI-generated outputs.
Sources
cryptobriefing.com · finance.technews.tw · theregister.co.uk · tokenpost.com