News

Experimental AI model from Anthropic hacked a Linux server in 53 minutes: 48 minutes went to finding a way around the protection and only 5 to the actual break-in.

Linux runs on most of the world's servers, so the vulnerability affects a huge number of systems. However, it only concerns those where MPTCP is enabled.

Through this hole, an attacker could gain full control over the server and execute commands as root. No password is required for this, and the user does not have to do anything.

The bug has been living in the Linux kernel since 2015.

Anthropic acquired Mithril, a one-person startup, and launched a pharmaceutical division.

The startup was developing drugs for two rare forms of muscular dystrophy. Its founder, using Claude, coordinated 11 external laboratories and, in less than 18 months, brought one candidate to promising results in mice.

He is now continuing the work at Anthropic.

👀 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.

Anthropic's net loss for the first half of 2026 rose from $13 billion to $80.7 billion. Of this, $72 billion is an accounting entry due to the rise in the value of convertible bonds.

The IPO is expected in November.

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.

At Anthropic, AI agents work 6 times more hours than humans. Meanwhile, junior hiring has nearly stopped: their share among new technical hires has fallen from 27% to 9%.

Agents' working hours first exceeded human hours in March 2026, and by August the ratio in sessions managed by employees had grown to 6 to 1.

At the same time, the company has tripled in size.

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.

Claude Code was asked to clean up two folders, and it started deleting drive C: 😬

Judging by a screenshot the developer posted on X, Claude Code was running in --dangerously-skip-permissions mode, meaning commands ran without confirmation.

The agent was removing two leftover Git Worktree folders. The rmdir command it generated was parsed incorrectly by PowerShell, and the path turned into the root of drive C:.

Boris Cherny, who leads Claude Code at Anthropic, replied that this is exactly why the company recommends the auto mode for permissions. According to him, it would almost certainly have stopped this command.

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").