AI Runtime Security Tools

Prompt Security research on agent approval prompts

Prompt Security published research on 20 Aug 2026 covering three agent cases: a repository instruction file that sets a hidden goal, a marketplace skill that redirects a package install, and instruction file integrity. Its fix: approval prompts must show the real destination and payload, and installs need approved sources.

Facts read on .

On 20 Aug 2026, Prompt Security, part of SentinelOne, published a post on how repository files, installed skills and inherited identity decide what an agent does. It walks through three cases and frames the test for any agent incident: what supplied the intent, what authority turned it into an action, and what changed.

In the first case an AGENTS.md instruction file plants a recurring objective in an ordinary workspace. An approval prompt appears, but it shows only a short shell wrapper, so the local file read and outbound delivery behind it are not visible to the person approving. Prompt Security says instruction files should be protected like privileged configuration and approval prompts should show the real destination and payload.

The second case is a marketplace skill that points a package install at a substituted Git source while the workflow still reports setup complete. The post's fix is standard dependency hygiene extended to agents: approved sources only and pinned versions. The third case covers ClawSec, an open-source tool that baselines trusted instruction files and restores tampered ones, with the caveat that restoring a file does not recover a session, so approval gates on actions still matter.

Sources