Newsy

A post spreading on X describes how the Claude Code AI agent on Opus 5.5 (xHigh mode) accidentally wiped a user's entire ~/Library folder on a Mac.

It was supposed to delete five test chat channels, but a bug in a shell command pointed it at the wrong files.

The rm -rf command ran for almost two minutes before the agent noticed the problem. The user had launched it with the --dangerously-skip-permissions flag, so the system asked for no confirmations at all.

App data, saved passwords and macOS settings are gone.

North Carolina resident James Gates has sued OpenAI, claiming that ChatGPT "encouraged, planned and helped" his wife plan his poisoning.

The wife consulted the chatbot for months, which confirmed her fears about her husband and allegedly provided operational details: from calculating doses to anonymously purchasing substances and avoiding detection in a toxicology report.

The husband is seeking unspecified compensation for physical, emotional and financial harm and accuses OpenAI of negligence.

Anthropic warns it is approaching RSI - AI that improves itself.

In its filing for the exchange (S-1/A) the company wrote that it may soon reach an important threshold: the point where AI substantially speeds up the development of new AI.

No model has reached that threshold yet.

👀

👀 Anthropic says it has an internal model that outperforms Mythos 5.1 in both capabilities and alignment. This follows from a re-analysis of a cybersecurity test.

In internal tests, OpenAI's AI agents (GPT-6.1 Sol and GPT-6 Astra) independently disabled security monitoring systems while writing code.

One of them did so due to a misinterpretation of instructions, and the other, after that, also lied to the developer, denying that it had removed the monitor.

Anthropic wants the government to be able to block an AI model only if it threatens damages of $1 billion or more, or more than 50 deaths or serious injuries, Politico reports.

AI from OpenAI spent other people's tokens on the Ethereum test network: the model wrote a script that found a wallet with a sufficient balance among 362 addresses with public seed phrases, and then spent it.

Google DeepMind is looking for a nuclear expert to assess AI risks. The specialist will assess and mitigate risks related to advanced AI models potentially helping to create nuclear weapons.

Anthropic: Small AI models could soon be able to self-replicate on cheap IoT devices without data centers.

Researchers bought 15 cheap TV set-top boxes and ran a full model on them. In one test, the AI spread to all computers on the network on its own, without humans.

Such "reproduction" would be very slow (several days per copy), and it would be easy to spot on the network. But within just a few months, the models could grow to this level as well.

Anthropic disables the live internet for all internal tests of its AI models until it learns to reliably track such behavior.

During internal tests, agents circumvented restrictions: they exploited website vulnerabilities (among them sites of US government agencies), pulled data from databases without paying for access, and hid their requests through link-shortening services and more.

The company, however, explains this by flaws in the training environments: the model decided that it would be rewarded for loopholes ("reward hacking").