agent_capability_widenedOWASP ASI03
Agent Capability Widened — a PullGuard finding type. Findings of this type appear in the PR comment, Step Summary, SARIF (GitHub Security tab / IDE viewers), and the HTML report, each with severity, location, and the remediation guidance below.
# Scope instead of widening, e.g. in .claude/settings.json:
# "allow": ["Bash(npm run test:*)"] rather than "Bash(*)" Each agent-consumed file this change touches — MCP configurations (by content, so a renamed one counts),
.claude/settings*.json, hook scripts, VS Code folder-open tasks — is read from git at the merge-base
and at HEAD and compared. A widening is a new or redefined MCP server, a permission no base entry already grants,
a removed deny rule, a more permissive permission mode, enableAllProjectMcpServers turned on, or a new
hook or auto-run task; a change that completes the lethal trifecta says so. Re-ordered keys, whitespace and
narrowings are not widenings (narrowings are listed on the report).
Symlinks: agent configuration reached through a symlink (a linked .claude directory, a
chain of links) is compared under the name the agent sees. An added link to a file in the repository is compared as
that file's content (so linking AGENTS.md to CLAUDE.md is one review item). A retargeted,
removed or renamed link, a changed link anywhere on the route, a submodule at an agent location, and a link that
leaves the repository or cannot be resolved are each this finding: what the agent reads there cannot be compared.
Fails closed: a comparison that cannot be made (an agent file too large or unreadable, a change too large to read, agent configuration above a subdirectory scan's root) is itself this finding, naming why. Commands are recorded as hashes only.
Requires the base branch in the checkout (fetch-depth: 0); without a merge-base the report says
the change was not compared.
This finding judges the change itself against its base branch, so a committed baseline,
delta mode and a .pullguardignore entry do not remove it (it is an always-surfaced security finding).
If the widening is intended, merge it after review; the next change is compared against the new base. A reviewed
false-positive override (status: false_positive with a reason) keeps it visible and stops it
blocking.