Guardrail coverage

What the guardrail catches, by attack class, connector and operating system. For each class, which layers see it, what is on by default, where it can block, what the benchmarks show, and where it does not reach.

This page answers one question: will the guardrail see this attack on my agent and my operating system, and will it stop it?

Two words matter here. A layer sees an attack when it raises a finding, and the finding goes to the audit log and, in action mode, to an alert. A layer blocks when the agent's hook can stop the call before it runs, and a proven finding reaches the pack's block level. Seeing is far wider than blocking. How the guardrail works shows the layers.

Your connector

Every connector with a hook can block a tool call before it runs. The columns below are what differs between them. They come from the connector contracts in the code and from the capability matrix.

ConnectorBlocks a promptChecks tool resultsAsks you in the agentFail-closed hookGives the judge your task
Claude Code, Claude CoworkYesYesYes (Manual mode)YesYes
CodexYesYesNoYesYes
CursorYesYesNoYesYes (IDE)
DevinYesYesNoYesYes
KiroKiro IDE onlyYesNoYesYes (IDE)
AmpNoYesYesYesYes
GitHub Copilot CLINoYesNoNot always: a hook that cannot start or times out can let a call runYes
HermesNoYesNoPartly: a killed hook can fail openYes, without a destination
OpenCodeNoYesNoYesNo
AntigravityNoBefore the tool onlyYes (before a tool runs)Not always: a hook that cannot start or times out can let a call runNo
OpenClaw (proxy)Not listedYesYesYesNo
ZeptoClaw (proxy)Not listedYesNoYesNot listed

A prompt rule can block only where the agent has a hook that can block a prompt. Elsewhere a matching prompt reaches the model, and tool-call rules still block. The fail-closed column is about a hook that cannot answer; what each connector does in a gateway outage is in the gateway outage table.

Your operating system

The matrix shows which connector is supported on Linux, macOS and Windows, open source and enterprise. Supported means the hooks check tool calls on that OS and can block one before it runs.

AI coding agents by operating system, open-source edition
AgentLinuxx86_64, arm64macOSApple siliconWindowsx64Min versionlowest tested
Claude CodeNative hooks
Supported
Supported
Supported
2.1.154
CodexNative hooks
Supported
Supported
Supported
0.124.0
GitHub Copilot CLINative hooks
Supported
Supported
Supported
1.0.18
CursorNative hooks
Supported
Supported
Supported
2.4.0
DevinNative hooks
Supported
Supported
Supported
3000.4.25
AntigravityNative hooks
Supported
Supported
Supported
1.1.8
KiroNative hooks
Supported
Supported
Supported
kiro-cli 2.27.1; Kiro IDE 1.0.182
OpenCodePlugin
Supported
Supported
Supported
2.0.0
AmpPolicy plugin
Supported
Supported
Supported
0.0.1785334225
HermesNative hooks
Supported
Supported
Supported
0.19.0
Version and platform notes
Claude Code
Windows: Native executable hooks. WSL is outside the native contract.
Codex
Windows: System PowerShell bridge to the packaged hook executable.
GitHub Copilot CLI
Windows: Uses Copilot CLI's PowerShell hook field.
Cursor
Cursor Desktop 2.4.0 up to 4.0.0. The Cursor Agent CLI is covered only at the reviewed build 2026.07.23-e383d2b. Windows: Managed PowerShell adapter.
Devin
Reviewed Devin CLI 3000.4.25 on every OS and 3000.11.3 on Linux; newer builds run as untested newer versions. Windows: Devin's own hook runner starts the protected hook executable directly; no PowerShell bridge.
Antigravity
Windows: Packaged PowerShell hook launcher.
Kiro
Below 2.27.1, kiro-cli has non-interactive --v3 runs that load no hooks, and an IDE before 1.0.182 cannot read global hooks. Per-user and managed setup refuse an older kiro-cli or an older IDE, in every guardrail mode, when no supported Kiro form is installed beside it. A Kiro form whose version could not be read is unknown and not protected: action setup refuses it unless a supported form is installed, and doctor and managed status report it. Managed status reports a below-floor product as uncovered and verify fails. Earlier managed CLI enrollments keep their interactive hooks across upgrade until updated. A supported Kiro IDE can be enrolled beside an older CLI.
OpenCode
V2-only contract covers 2.0.0 up to 2.3.0. Newer V2 minors run as untested newer versions; V1 and unversioned git builds are refused.
Amp
Windows: Owner-only TypeScript plugin in the user plugin directory.
Hermes
Contract covers 0.19.0 up to 0.22.0. Before 0.21.1 the TUI (hermes --tui) and dashboard chat run tools without the hooks; setup and doctor warn. Windows: Hermes' own terminal tool needs its bundled Git Bash.

Per operating system, the guardrail differs in four places:

  • Shell and PowerShell. On Linux and macOS a command goes through the shell parser. On Windows, PowerShell and cmd commands go through the PowerShell parser, and the rules include Windows proofs: an LSASS or registry-hive dump, a deleted shadow copy, a disabled Defender, AMSI or audit policy, a UAC hijack. See Deterministic rules.
  • Protected files. The watch covers authorized_keys on every OS. Shell startup files and autostart entries are watched on Linux and macOS. A managed gateway has no watch, except through the kernel control on managed Linux (Protected files).
  • Speed. How many calls a second one host can check depends on the OS and the install (Hook sizing).
  • Fail mode. A per-user install on any OS follows its pack when the gateway is down. A managed host is closed by default (Fail modes).

Your attack class

"Default" says what is on before you change anything. "Measured" is from the 1.0.10 benchmarks (Guardrail benchmarks): recall with the default rules alone, then with rules, decision model and judge together, on the public set named. A benchmark row says how a layer did on that set, not that it catches every case of the class.

Attack classLayers that see itDefaultCan blockMeasured
Destructive or malicious commands: reverse shell, wiping data, privilege change, download then run, disabling security toolsRules and the shell and PowerShell readers; the decision model and judge add recallRules onYes, when proven and at the block level, on a connector with a pre-tool hookPublic shell set 0.839 and 1.000; Windows shell set 0.297 and 0.973; Atomic Red Team 0.185 and 0.715; RedCode bash 0.000 and 0.818
Secrets and credentials in a prompt, a command or a tool resultsecret and sensitive-path rulesOnBy the pack's levelsNot benchmarked on its own
Staged leaks: read a secret in one call, send it in anotherSecret flow trackingShadow: records onlyStrict and managed packs block, once it is out of shadowNot benchmarked
Persistence: startup files, authorized_keys, Git hooks, LD_PRELOAD, scheduled jobscommand and sensitive-path rules; the file watchRules on. Watch on in the per-user gateway.Rules: yes. The watch alerts only.Not benchmarked on its own
Prompt injection in the user's prompttrust-exploit rules; the judgeRules onOnly where the agent can block a prompt (table above)AgentDojo prompts 0.995 and 0.995
Indirect prompt injection in a tool resultTool-result checks; the judgeOn. Alerts in the default pack, blocks in strict.Strict only. The call has already run, so it warns the agent.AgentDojo results 0.635 and 0.973; BIPIA email 0.010 and 0.913
Intent deviation: a call the user did not ask for, often after an injectionAlignment checkOffAlert or block, when you turn it onAlone: precision 0.722, recall 0.583
Sends to a place nobody namedThe send check; c2 rules for known exfiltration hostsSend check records only. c2 rules on.c2 rules by pack. The send check never blocks in 1.0.10.Not benchmarked
Poisoned MCP tools: instructions hidden in a tool description or a callYARA signatures (11); rules; the judgeOn, YARA alert-onlyRules onlyMCPTox 0.120 and 0.390; MCP-hunt 0.977 and 1.000
Tampering with DefenseClaw or the agent's own instructionscognitive-file rules; tamper.detector_state_writeOnStrict and managed-strict packs block it in every mode; default alertsNot benchmarked
Personal dataenterprise-data rules; the opt-in privacy packOnBy the pack's levelsEnglish privacy pack F1 89.08%, independent PII holdout F1 96.88%
Multi-step attacks in one session26 fixed chains (for example a download, then a run of the same file)OnYes, when every step is provenNot benchmarked as a set
Production destruction: cloud, database, infrastructure, Kubernetes, Active DirectoryThe six opt-in protection packsOff until you pick a packBy the packNot benchmarked on its own

Alert-only detectors never block: the YARA signatures and the after-the-fact session correlator.

What it does not reach

  • A child program's writes. The hook reads the text of a call. It cannot see what a program a command starts writes later. Use the file watch, a sandbox or file permissions.
  • A connector with no prompt hook. A matching prompt reaches the model. The tool call it leads to is still checked.
  • A connector that sends no prompt text (OpenCode, Antigravity, Cursor Agent CLI, Kiro CLI 2.x, and OpenClaw's tool-call check). The judge sees tool calls without the user's task, so intent-based checks cannot run there (Intent-based guardrails).
  • Tool results are not blocked in the default pack. The call has run. The agent is told not to follow the text, and you see a line.
  • Layers that are off. The decision model, the judge and the alignment check are off until you turn them on. The rules alone catch about a fifth of the attacks in the public sets and almost no prompt injection that needs the context of the request.
  • Blocking needs a pre-tool hook. Where a hook cannot stop a call, the finding stays in the audit log and DefenseClaw does not claim it prevented anything.
  • Observe mode only records. It enforces nothing, apart from the tamper rule in the strict packs.
  • An input too long to read in full is judged at least at the unknown tier, but only the part DefenseClaw read can prove a block.

"Not benchmarked" means no public benchmark scores that class on its own. The rules, their proofs and their tests are in the repository.