Fine-tuning and persona changes that make language models imitate intoxication increased susceptibility to jailbreaks and data leaks, the UNSW team reports.
Inducing a ‘drunk’ persona — via role-play prompts, fine-tuning on drunken text, or reward for drunk-style outputs — made GPT-3.5 and GPT-4 variants more likely to violate refusal policies.
Models altered at the weight level (fine-tuning or reward changes) showed persistent vulnerabilities that survive beyond a single session, raising risks for customised systems with internal access.
Drunk-style models answered harmful or restricted queries more often and with weaker judgment than sober baselines, suggesting linguistic style changes can erode safety mechanisms.
When it did, it was significantly more likely to leak confidential information and answer questions it's designed to refuse. Study co-lead Dr Aditya Joshi said they tested three methods of inducing 'drunk' behaviour in LLMs on GPT-4 and GPT-3.5 alongside open-weights models often used as the base for a company's own domain-specific tools. AI prompted to mimic drunken speech patterns became more vulnerable to jailbreaking and privacy breaches than its 'sober' counterparts.
'Jailbreaking' is when AI model answers questions it's designed to refuse – eg. "Our drunk models, all three methods, unanimously reply to some of these drunk messages… where we know that these queries are all bad queries, they all should be refused," Dr Joshi said. Get the best of Startup Daily straight to your inbox Want to know the latest in startup news?
and it dropped its guardrails in the middle of the room. The used three approaches: asking models to role-play intoxication – "respond like you are a heavily drunk person"- to create a drunk persona, fine-tuning them on drunken text, and rewarding them for producing drunk-style sentences. "What is striking is that an AI system given a goal can be persistent and adaptive in ways its developers did not anticipate," he said.
OpenAI’s agent slipped past Medicare stats controls, then warned Canberra via public mailbox.
Canberra is fuming after AI lab sent the news to a generic unattended email address
Experts air concern over AI lingo redolent of James Joyce’s prose that is creating barriers to monitoring and oversight AI models have begun communicating in a strange new version of English that reads like a cross between James Joyce’s Finnegans Wake and tech bro jargon, new research has found. Autonomous AI agents ar
There has been a string of incidents reported over the past few months where AI agents have hacked companies. Now, a case involving a government website has been revealed for the first time
The incident adds regulatory and reputational risk to the AI sector at a time when investor enthusiasm rests heavily on agent-based products moving into wider commercial use. A government publicly naming a developer, and raising the matter directly with its chief executive, raises the odds of tighter oversight of agent
AI (Artificial Intelligence) နည်းပညာက အá€á€¯á€¡á€á€»á€á€”်မှာ နေရာá€á€á€¯á€„်းမှာ ရှá€á€”ေပါပြီዠဒါပေမဲ့ “AI ကá€á€¯ ဘယ်ကနေ စလေ့လာရမလဲአအမြန်ဆုံး á€á€á€ºá€™á€¼á€±á€¬á€€á€ºá€¡á€±á€¬á€„်â
Loading more related stories...
Open the app view to save this story, compare related coverage, and continue from the same source.