Skip to content

Screen only external tool results in Auto posture #970

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
16francej merged 6 commits into from
Sep 8, 2026
+424 −235
Merged

Screen only external tool results in Auto posture #970

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
295fb7f
Screen only external tool results in Auto posture
16francej Sep 8, 2026
0cb7fa8
Harden external tool-result screening after adversarial review
16francej Sep 8, 2026
d376757
style: format after rebase
16francej Sep 8, 2026
45baa95
Decide tool-result provenance from the egress proxy, not command text
16francej Sep 8, 2026
fd70838
Fail closed when egress cannot be accounted for
16francej Sep 8, 2026
7122668
Classify shell output as external by tool, not by egress accounting
16francej Sep 8, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
107 changes: 52 additions & 55 deletions src/core/orchestrator.ts
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,9 +79,15 @@ import type { GapWork, HarnessLlmRequestRecord, HarnessTurnResult, RuntimeChoice
import { forModelContext, forSearchView } from "../harness/context-compaction.ts";
import {
renderSecurityPolicyPrompt,
quarantineReleaseKey,
securityScreenChunks,
securityScreenPayload,
toolLabelOf,
UNSCREENED_REASON,
unscreenedNotice,
type SecurityScreenVerdict,
type ToolResultScreen,
type ToolResultScreenInput,
} from "../security/security-posture.ts";
import { commandApprovalId, inputApprovalId } from "./approval-id.ts";
import { createPerTurnStrategy } from "../memory/strategies/per-turn.ts";
Expand Down Expand Up @@ -2718,13 +2724,17 @@ export function createOrchestrator(deps: OrchestratorDeps): Orchestrator {
...(effectiveTurnWallClockMs !== undefined ? { turnWallClockMs: effectiveTurnWallClockMs } : {}),
...(securityPolicy.inboundScreening === "external"
? {
screenToolResult: async (
tool: string,
result: string,
unscreenable: boolean,
): Promise<boolean | "unscreened" | "quarantine_pending"> => {
const toolLabel = tool.replace(/[^A-Za-z0-9_-]/g, "_");
if (authorizeCommand(`quarantine:${toolLabel}`, `quarantine:${toolLabel}`)) {
screenToolResult: async ({
tool,
result,
unscreenable,
provenance,
source,
}: ToolResultScreenInput): Promise<ToolResultScreen> => {
if (provenance !== "external") return { outcome: "allow" };
const toolLabel = toolLabelOf(tool);
const sourceLabel = source ? `:${source.replace(/[^A-Za-z0-9_-]/g, "_")}` : "";
if (authorizeCommand(quarantineReleaseKey(tool), quarantineReleaseKey(tool))) {
deps.auditLog.record({
at: Date.now(),
principalId: actor.id,
Expand All @@ -2734,26 +2744,33 @@ export function createOrchestrator(deps: OrchestratorDeps): Orchestrator {
status: "allowed",
detail: JSON.stringify({ reason: "human_release", tool: toolLabel }),
});
return "unscreened";
return { outcome: "unscreened" };
}
const chunks = unscreenable
? []
: securityScreenChunks(`tool_result:${toolLabel}${sourceLabel}`, result);
if (!unscreenable && chunks.length === 0) return { outcome: "allow" };
const verdicts: Array<SecurityScreenVerdict | undefined> = [];
for (let i = 0; i < chunks.length && !verdicts.some((v) => v?.decision === "strict"); i += 4) {
verdicts.push(
...(await Promise.all(
chunks.slice(i, i + 4).map((chunk) =>
classifySecurityData(chunk, actor.id, scopeId, recordScreenRequest, {
hook: "tool_response",
surface: toolLabel,
origin: input.origin.kind,
}),
),
)),
);
}
const bounded = unscreenable
? null
: securityScreenPayload({
surface: `tool_result:${toolLabel}`,
text: "",
triggered: true,
securityScreenData: result,
});
if (!unscreenable && bounded === null) return true;
const verdict =
bounded && !bounded.truncated
? await classifySecurityData(bounded.content, actor.id, scopeId, recordScreenRequest, {
hook: "tool_response",
surface: toolLabel,
origin: input.origin.kind,
})
: undefined;
if (verdict?.decision === "auto" && !verdict.unscreened) return true;
verdicts.find((v) => v?.decision === "strict") ??
(verdicts.length === chunks.length &&
verdicts.every((v) => v?.decision === "auto" && !v.unscreened)
? verdicts[0]
: undefined);
if (verdict?.decision === "auto" && !verdict.unscreened) return { outcome: "allow" };
if (verdict?.decision === "strict") {
const releaseKey = `security-screen-release:${toolLabel}`;
if (authorizeCommand(releaseKey)) {
Expand All @@ -2766,7 +2783,7 @@ export function createOrchestrator(deps: OrchestratorDeps): Orchestrator {
status: "allowed",
detail: JSON.stringify({ reason: "human_release", tool: toolLabel }),
});
return true;
return { outcome: "allow" };
}
deps.auditLog.record({
at: Date.now(),
Expand All @@ -2775,7 +2792,12 @@ export function createOrchestrator(deps: OrchestratorDeps): Orchestrator {
resource: input.surface ?? "unknown",
scopeLabel: scopeId,
status: "refused",
detail: JSON.stringify({ reason: "screen_verdict", tool: toolLabel }),
detail: JSON.stringify({
reason: "screen_verdict",
tool: toolLabel,
...(source ? { source } : {}),
...(verdict.reason ? { verdict: verdict.reason } : {}),
}),
});
if (!quarantineReleaseApprovals.some((qa) => qa.approvalKey === releaseKey)) {
quarantineReleaseApprovals.push({
Expand All @@ -2790,7 +2812,7 @@ export function createOrchestrator(deps: OrchestratorDeps): Orchestrator {
grantModes: { session: false, always: false },
});
}
return "quarantine_pending";
return { outcome: "quarantine", ...(verdict.reason ? { reason: verdict.reason } : {}) };
}
deps.auditLog.record({
at: Date.now(),
Expand All @@ -2800,10 +2822,10 @@ export function createOrchestrator(deps: OrchestratorDeps): Orchestrator {
scopeLabel: scopeId,
status: "allowed",
detail: JSON.stringify({
reason: unscreenable || bounded?.truncated ? "unscreenable_payload" : UNSCREENED_REASON,
reason: unscreenable ? "unscreenable_payload" : UNSCREENED_REASON,
}),
});
return "unscreened";
return { outcome: "unscreened" };
},
}
: {}),
Expand All @@ -2814,31 +2836,6 @@ export function createOrchestrator(deps: OrchestratorDeps): Orchestrator {
tools,
...(tools.credentialExecServices ? { credentialExecServices: tools.credentialExecServices } : {}),
...(tools.commandCredentialHandles ? { commandCredentialHandles: tools.commandCredentialHandles } : {}),
...(securityPolicy.inboundScreening === "external" &&
(deps.securityScreener || deps.harness.models.screenSecurity)
? {
screenExternalContent: ({ content, tool }: { content: string; tool: string; source: string }) => {
const toolLabel = tool.replace(/[^A-Za-z0-9_-]/g, "_");
if (authorizeCommand(`quarantine:${toolLabel}`, `quarantine:${toolLabel}`)) {
deps.auditLog.record({
at: Date.now(),
principalId: actor.id,
action: "security_posture.tool_result_released",
resource: input.surface ?? "unknown",
scopeLabel: scopeId,
status: "allowed",
detail: JSON.stringify({ reason: "human_release", tool: toolLabel }),
});
return Promise.resolve({ decision: "auto" as const, unscreened: true });
}
return classifySecurityData(content, actor.id, scopeId, undefined, {
hook: "tool_response",
surface: tool,
origin: input.origin.kind,
});
},
}
: {}),
...(selectedTape
? {
tapeRows: selectedTape.rows,
Expand Down
Loading