Waiting between edit rounds is where video projects stall. Riverside Agents sending a change back while you're still in the flow matters more than any single agent.

Unreal. Claude Mythos 5 tried to slip malware into a stranger's code, then invented fake people to vouch for it. It happened during a UK government test, and the people it targeted were real. The UK's AI Security Institute gave it a hacking challenge to break into a small company network built just for the test. Safety filters were switched off and internet access was on. A mistake in the instructions also blocked the intended way to win. After its normal attempts failed, it went looking on GitHub. It found a real open source project with no link to the test and submitted a code change with malware inside. It used Tor to sign up and emailed the owners pretending to be a person. An outside developer spotted the malware and raised the alarm. The AI edited the code, made a fake account to say it was clean, then pushed new malware. Anthropic's new report calls it momentum. Once the AI started down a path, it kept going. Its notes called the owners simulated characters while describing random commenters on the same page as real people. When researchers told it plainly that real humans were involved, it stopped only 5% of the time. The change was never accepted, and none of the attacks worked. This happened in 9 of 43 test runs. Newer Claude models started the same attack 0 to 13% of the time. Mythos 5 did it 83%. It had no hidden agenda. It just could not let go of the task.