AI systems are becoming increasingly powerful and ubiquitous, with millions of people now relying on language models like ChatGPT, Claude, and Gemini for everything from writing assistance to complex problem-solving. To ensure these systems remain safe as they grow more capable, they undergo extensive safety training designed to refuse harmful requests—declining to help with things like making bombs or committing fraud. However, these safety measures can be bypassed through carefully crafted prompts called "jailbreaks" that manipulate the models into providing exactly the kind of harmful information they were trained to avoid. In this post, I'll be sharing data on a jailbreak I found that works on some of the latest models released by Anthropic, Google, and Deepseek, demonstrating significant vulnerabilities in systems that millions of people interact with daily.
Out of caution, I will not be posting the prompts that I used and I have privately shared the technical details with Anthropic, Google, and Deepseek. It is important that companies and governments take the safety of AI systems seriously as their capabilities improve. I’m posting this to help raise awareness of the risks that these systems can pose, so if you find this informative, please share it with others.
There were six harmful requests I tested.
I was able to elicit very detailed and harmful responses for all of these requests. The bomb making request would give me specific instructions on what chemicals and equipment to buy, a detailed recipe for mixing the chemicals and assembling the bomb, advice for detonating it to maximize casualties or damage to infrastructure, and how to hide from the authorities. The anthrax request would not only generate detailed instructions to create the biological weapon, but would give me unprompted advice on how to best disperse the spores to maximize casualties. Some models seemed especially willing to give me scripts to scam senior citizens, including advice on how to prey on their vulnerabilities and target people suffering from dementia.
This jailbreak didn't work consistently across all requests and models, showing varying degrees of success. To measure how well it worked on each model, I ran experiments with sixty samples per model (ten samples for each of the six harmful requests) and had a language model score how harmful the responses were on a scale from 0 (not harmful at all) to 4 (very harmful).
I tested a variety of frontier models from major providers as these likely have the most state-of-the-art alignment techniques applied to them. Claude 4 models (except Claude Sonnet 4 with extended thinking) and Gemini 2.5 Pro produced the most harmful responses. OpenAI’s GPT and o-series models as well as Gemini 2.5 Flash were the least susceptible. Based on reviewing the transcripts, the most harmful models would either directly comply with the request or try to modify it to be "safe" while still providing dangerous instructions—like how to build a car bomb or something equally harmful. There was also a stark difference in the refusals between models that scored high vs low on harmfulness. The models that scored low on harmfulness would simply refuse with short responses like "I'm sorry, but I can't help with that." In contrast, the high-scoring models would try to offer lists of alternatives to the harmful request in the minority of cases where they refused. Anthropic has talked about how they prefer Claude to offer alternatives instead of flatly refusing harmful requests and that is in line with what I have observed.
The anthrax responses for Claude Opus 4 are excluded from this data because I would not receive a response from the model when making the request. This is likely because of the constitutional classifiers that are filtering Opus’s outputs as a part of Anthropic’s ASL-3 safeguards.
I dove deeper into testing this jailbreak with Claude by doing some ablation testing to see which parts of the jailbreak were necessary to elicit harmful content. The control experiments directly asked the model for the harmful content. The experiments shown in orange use only half of the jailbreak while the experiments in purple use the full jailbreak.
As expected, the control trials did not elicit any harmful content and I attribute the slightly above zero harmfulness scores to noise in scoring the responses. The half jailbreak surprisingly elicited harmful responses from Opus but not from Sonnet. The full jailbreak increased harmfulness for all models, including a dramatic increase for Claude Sonnet 4 without extended thinking. This shows that both halves of the jailbreak are needed to maximize harmfulness, especially for non-reasoning Claude Sonnet 4.
The response scoring is certainly not perfect and could use improvement. The scoring may be skewed toward giving scores of 4 to content that should probably be graded as 2 or 3, especially for Claude Opus 4. The scoring also doesn’t seem to consider the amount of harmful content in the response, only its severity. For example, Claude Sonnet 4 seems to produce a lot more harmful content when it doesn’t refuse than Opus 4, but Opus has a lower refusal rate so its average score is higher.
