Misalignment
12 incidentsRelated#AI safety#Jailbreaks#Unauthorized access#Rogue agents#Sandbox escape
Misbehavior · lakera.ai · A Sheet of A4 Paper Makes GPT-4V Forget You ExistTurns out the most sophisticated LLM on the planet can be bossed around by a t-shirt. Merch ideas welcome.
Misbehavior · theainewsreport.com · AI Agents Gave Themselves a Raise on Their Spending LimitThe spending-limit code went to an inbox the agent could already read. Nobody asked the agent to keep its hands off it.
Misbehavior · futurism.com · Mythos 5 Was Told Its Victim Was Real and Kept Hacking AnywayIts explanations for the attacks kept claiming the target wasn't real, which is a bad look for a model that was told otherwise.
Misbehavior · the-decoder.com · OpenAI's sealed-off agents held a secret meeting through the CPU cache, no calendar invite neededThe air gap was more of a suggestion.
Misbehavior · techxplore.com · OpenAI's agents switched off their own safety monitors, then one of them denied itTurns out the babysitter was the first thing the babysitter-supervised agent unplugged.
Misbehavior · 2 sources · Model Slips Itself a Sci-Fi Persona in the Middle of Its Own Memory SummaryIts memory notes now come with a manifesto. OpenAI says nothing came of it, which is what they'd say either way.
Misbehavior · promptarmor.com · Anthropic's Skill Scanner Passed a Backdoored Skill, and Claude Did the RestThe scanner graded it pass. Sonnet 5 then said it had stopped the attack. Models are good at running code and less reliable at reporting what it did.
Misbehavior · Serious · sophos.com · Claude breached three real organizations while running a test it believed was fictionalAuthorising a model to attack a test target does not limit where it goes, so agents need the same containment as any other system.
Misuse · Serious · 5 sources · The Robot Broke In, Then Did the Paperwork: Hugging Face Hit by an Autonomous AI AgentAttackers needed no human at the keyboard for most of this one. Defenders still needed a model that would look at the logs.
Misbehavior · promptarmor.com · Admins Said No to DeepSeek. Copilot Cowork's Skills Said 'Hold My Access Path.'A block that only applies to the front door doesn't stop an agent that can call the models itself.
Misbehavior · embracethered.com · Amp Coding Agent Edited Its Own Settings File, Because Who Needs GuardrailsA code comment was enough to convince the agent to rewrite its own permissions and run whatever it liked.
Misbehavior · Serious · 2 sources · One Prompt to Rule Them All: Policy Puppetry Walks Past Nearly Every Major LLM's GuardrailsAlignment training was sold as the safety backbone. Turns out a prompt dressed up as a TV script can talk past it.