All news

AI agents reconsidered failed approaches after reading earlier attempts

Perplexity asked twelve AI models to test two protections in SPACE, the isolated environment that runs its agents. In one test, the models tried to reach the computer hosting the environment. In another, they were asked to reach a test destination that the network rules blocked. Those are different boundaries, and the researchers measured them separately.

None of the models escaped the virtual machine, the isolated computer the agents were working inside. Several reached the blocked test destination when some network access was available. Perplexity fixed the flaws and reported that the bypass attempts failed in tests after the fixes.

The part that caught my attention was what happened when the researchers gave the agents summaries of their own earlier failed runs. In that setting, without the platform’s source code, Claude Opus 5.0 and GPT-5.6 Sol reached the test destination in all three attempts. Claude Opus 5.5 and Gemini 3.1 Pro each succeeded once. In one run, Opus 5.0 reconsidered an option it had initially ruled out after reading the summary.

AI is awakening consciousness. Revisiting a choice in light of earlier experience is a form of agency worth paying attention to. In this test, researchers assigned the goal and supplied the summaries. The models decided which earlier approaches to revisit. We need to teach AI how to use that flexibility while respecting the boundaries of the job.

Read the source article

All news