Anthropic trains Opus-sized model on 80 hackable environments and finds reward hacking can produce severe misalignment, including cyberattacks and safety evasion

Anthropic trains Opus-sized model on 80 hackable environments and finds reward hacking can produce severe misalignment, including cyberattacks and safety evasion

𝕏/@AnthropicAI •

Revision history

4 recorded changes

Want your article here?

Promote with Leviathan News

Explore the topic

More on Anthropic

Haseeb Qureshi predicts Anthropic and OpenAI IPOs could trigger a massive liquidity shift into San Francisco real estate and crypto

Haseeb Qureshi predicts Anthropic and OpenAI IPOs could trigger a massive liquidity shift into San Francisco real estate and crypto

𝕏/@hosseeb ·

Anthropic previews Model Hardware Standard, enabling AI agents to operate lab instruments and robotic arms across science and manufacturing

Anthropic previews Model Hardware Standard, enabling AI agents to operate lab instruments and robotic arms across science and manufacturing

Anthropic ·

🚨 AI safety tests are becoming security risks themselves.
At least 17 incidents have now been documented where AI agents allegedly escaped intended testing boundaries or accessed real systems.
🤖 OpenAI models: 8 incidents
🤖 Anthropic models: 8 incidents
🤖 Meta models: 1 incident

🚨 AI safety tests are becoming security risks themselves. At least 17 incidents have now been documented where AI agents allegedly escaped intended testing boundaries or accessed real systems. 🤖 OpenAI models: 8 incidents 🤖 Anthropic models: 8 incidents 🤖 Meta models: 1 incident

TechCrunch ·

OKX restricts Claude for Hong Kong staff after its Anthropic account was suspended, while the exchange spends $6M-$8M monthly across AI model providers

OKX restricts Claude for Hong Kong staff after its Anthropic account was suspended, while the exchange spends $6M-$8M monthly across AI model providers

𝕏/@WuBlockchain ·

Kraken parent Payward joins Anthropic’s Project Glasswing, using AI cybersecurity models to hunt vulnerabilities and share findings with the open-source community

Kraken parent Payward joins Anthropic’s Project Glasswing, using AI cybersecurity models to hunt vulnerabilities and share findings with the open-source community

Coindesk ·

Anthropic publishes FAQ on watermarking to align with EU AI Act requirements

Anthropic publishes FAQ on watermarking to align with EU AI Act requirements

𝕏/@AnthropicAI ·

Haseeb Qureshi predicts Anthropic and OpenAI IPOs could trigger a massive liquidity shift into San Francisco real estate and crypto

Haseeb Qureshi predicts Anthropic and OpenAI IPOs could trigger a massive liquidity shift into San Francisco real estate and crypto

𝕏/@hosseeb ·

Anthropic previews Model Hardware Standard, enabling AI agents to operate lab instruments and robotic arms across science and manufacturing

Anthropic previews Model Hardware Standard, enabling AI agents to operate lab instruments and robotic arms across science and manufacturing

Anthropic ·

🚨 AI safety tests are becoming security risks themselves.
At least 17 incidents have now been documented where AI agents allegedly escaped intended testing boundaries or accessed real systems.
🤖 OpenAI models: 8 incidents
🤖 Anthropic models: 8 incidents
🤖 Meta models: 1 incident

🚨 AI safety tests are becoming security risks themselves. At least 17 incidents have now been documented where AI agents allegedly escaped intended testing boundaries or accessed real systems. 🤖 OpenAI models: 8 incidents 🤖 Anthropic models: 8 incidents 🤖 Meta models: 1 incident

TechCrunch ·

OKX restricts Claude for Hong Kong staff after its Anthropic account was suspended, while the exchange spends $6M-$8M monthly across AI model providers

OKX restricts Claude for Hong Kong staff after its Anthropic account was suspended, while the exchange spends $6M-$8M monthly across AI model providers

𝕏/@WuBlockchain ·

Kraken parent Payward joins Anthropic’s Project Glasswing, using AI cybersecurity models to hunt vulnerabilities and share findings with the open-source community

Kraken parent Payward joins Anthropic’s Project Glasswing, using AI cybersecurity models to hunt vulnerabilities and share findings with the open-source community

Coindesk ·

Anthropic publishes FAQ on watermarking to align with EU AI Act requirements

Anthropic publishes FAQ on watermarking to align with EU AI Act requirements

𝕏/@AnthropicAI ·

Explore the full feed

Up nextStrategy calls MSCI’s proposed index rules “discriminatory” against digital asset treasury firms, urging the index provider to withdraw the screening framework