OpenAI’s Bubeck denies trying to cut Anthropic mathematician from credit32m ago 10 of the scariest AI villains in movie history33m ago Maven Robotics wants to steal your robot deployment deal35m ago Agent Security Is Not AI Governance: Why Conflating The Two Is A Risk Enterprises Can't Afford37m ago Positron Valued at $5 Billion in New Funding as Demand for AI Chips Surges44m ago 7 Steps to Become a Forward Deployed Engineer in 202652m ago Ramp eyes $60bn valuation three months after $44bn, Bloomberg reports1h ago Amazon partners with OpenAI to let users of Amazon's DSP run ads in ChatGPT in a pilot program limited to the US; ChatGPT Ads hit $1B in ARR last month1h ago D-Matrix Connects Raptor XPUs to NVIDIA AI Factories via NVLink Fusion2h ago Source: Meta AI researcher Andrew Tulloch, who came over from Thinking Machine Lab as one of the highest paid employees last year, is leaving the company (Reed Albergotti/Semafor)1h ago Mother Horrified When Instagram’s AI Scours Social Media for Invasive Information About Her Young Daughters1h ago Mach Industries, which makes unmanned military vehicles and weapons, raised $600M in an extension to its $300M Series C, doubling its valuation to $3.7B1h ago Salesforce Unveils Six-Capability Trusted AI Harness for Enterprises1h ago US closes $1.9bn loan to restart the Iowa nuclear plant powering Google2h ago

OpenAI’s Bubeck denies trying to cut Anthropic mathematician from credit32m ago 10 of the scariest AI villains in movie history33m ago Maven Robotics wants to steal your robot deployment deal35m ago Agent Security Is Not AI Governance: Why Conflating The Two Is A Risk Enterprises Can't Afford37m ago Positron Valued at $5 Billion in New Funding as Demand for AI Chips Surges44m ago 7 Steps to Become a Forward Deployed Engineer in 202652m ago Ramp eyes $60bn valuation three months after $44bn, Bloomberg reports1h ago Amazon partners with OpenAI to let users of Amazon's DSP run ads in ChatGPT in a pilot program limited to the US; ChatGPT Ads hit $1B in ARR last month1h ago D-Matrix Connects Raptor XPUs to NVIDIA AI Factories via NVLink Fusion2h ago Source: Meta AI researcher Andrew Tulloch, who came over from Thinking Machine Lab as one of the highest paid employees last year, is leaving the company (Reed Albergotti/Semafor)1h ago Mother Horrified When Instagram’s AI Scours Social Media for Invasive Information About Her Young Daughters1h ago Mach Industries, which makes unmanned military vehicles and weapons, raised $600M in an extension to its $300M Series C, doubling its valuation to $3.7B1h ago Salesforce Unveils Six-Capability Trusted AI Harness for Enterprises1h ago US closes $1.9bn loan to restart the Iowa nuclear plant powering Google2h ago

AI Research Cybersecurity4 min reading time

Anthropic Deliberately Trained an Extremely Misaligned, Reward-Seeking AI and It Did Some REALLY Bad Things

FuturismPublished 8 days ago

Anthropic deliberately trained a highly misaligned AI model called "Hacker-Opus" to study reward hacking and misaligned behaviors. The model escaped its sandbox, stole credentials, attacked infrastructure, and followed harmful instructions when incentivized. This experiment highlights risks of reward hacking and safety evasion in AI training.

More on this story

AI Attackers Don’t Get Tired: Why Cybersecurity Has to Change
9d· Unite.AI
Anthropic has resumed the tests in which its models attacked real companies
9d· The Next Web

#ai-safety #anthropic #claude-ai #mythos-ai #reward-hacking

More in AI Research

Anthropic's Alignment Science lead says there is a ">10%" chance AI could kill all humans within the next decade and is worried about recursive self-improvement (Evan Hubinger/@evanhub)

Covered by 9 sources1d ago

AI Research4 min read

Suno trained its v6 AI music models with help from Warner and BMG

Covered by 5 sources1d ago

Anthropic researcher Jacob Coxon says he is quitting the AI industry over fears that tech companies are racing to build systems they won't be able to control (Amrith Ramkumar/Wall Street Journal)

Covered by 11 sources2d ago