diff --git a/src/agent/branch.rs b/src/agent/branch.rs
index f9554342a..001eab7ea 100644
--- a/src/agent/branch.rs
+++ b/src/agent/branch.rs
@@ -154,11 +154,14 @@ impl Branch {
 
         // Scrub tool secret values from the conclusion before sending to the
         // channel. Branches can spawn workers whose output may contain secrets.
+        // Layer 1: exact-match redaction of known secrets from the store.
+        // Layer 2: regex-based redaction of unknown secret patterns.
         let conclusion = if let Some(store) = self.deps.runtime_config.secrets.load().as_ref() {
             crate::secrets::scrub::scrub_with_store(&conclusion, store)
         } else {
             conclusion
         };
+        let conclusion = crate::secrets::scrub::scrub_leaks(&conclusion);
 
         // Send conclusion back to the channel
         let _ = self.deps.event_tx.send(ProcessEvent::BranchResult {
diff --git a/src/agent/channel_dispatch.rs b/src/agent/channel_dispatch.rs
index b86fd7feb..d406271dd 100644
--- a/src/agent/channel_dispatch.rs
+++ b/src/agent/channel_dispatch.rs
@@ -249,6 +249,7 @@ async fn spawn_branch(
     let event_tx = state.deps.event_tx.clone();
     let agent_id = state.deps.agent_id.clone();
     let channel_id = state.channel_id.clone();
+    let secrets_snapshot = state.deps.runtime_config.secrets.load().clone();
 
     let branch_span = tracing::info_span!(
         "branch.run",
@@ -260,11 +261,22 @@ async fn spawn_branch(
         async move {
             if let Err(error) = branch.run(&prompt).await {
                 tracing::error!(branch_id = %branch_id, %error, "branch failed");
+                // Scrub the failure message in case the error contains secrets
+                // (e.g. from failed tool calls echoing back prompt content).
+                // Layer 1: exact-match redaction of known secrets from the store.
+                // Layer 2: regex-based redaction of unknown secret patterns.
+                let raw = format!("Branch failed: {error}");
+                let conclusion = if let Some(store) = secrets_snapshot.as_ref() {
+                    crate::secrets::scrub::scrub_with_store(&raw, store)
+                } else {
+                    raw
+                };
+                let conclusion = crate::secrets::scrub::scrub_leaks(&conclusion);
                 let _ = event_tx.send(crate::ProcessEvent::BranchResult {
                     agent_id,
                     branch_id,
                     channel_id,
-                    conclusion: format!("Branch failed: {error}"),
+                    conclusion,
                 });
             }
         }
@@ -591,11 +603,14 @@ where
             Ok(Ok(text)) => {
                 // Scrub tool secret values from the result before it reaches
                 // the channel. The channel never sees raw secret values.
+                // Layer 1: exact-match redaction of known secrets from the store.
+                // Layer 2: regex-based redaction of unknown secret patterns.
                 let scrubbed = if let Some(store) = &secrets_store {
                     crate::secrets::scrub::scrub_with_store(&text, store)
                 } else {
                     text
                 };
+                let scrubbed = crate::secrets::scrub::scrub_leaks(&scrubbed);
                 Ok(scrubbed)
             }
             Ok(Err(error)) => {
@@ -608,6 +623,7 @@ where
                         } else {
                             message
                         };
+                        let scrubbed = crate::secrets::scrub::scrub_leaks(&scrubbed);
                         Err(WorkerCompletionError::Failed { message: scrubbed })
                     }
                 }
diff --git a/src/agent/cortex.rs b/src/agent/cortex.rs
index d2ff4811b..4fca18133 100644
--- a/src/agent/cortex.rs
+++ b/src/agent/cortex.rs
@@ -1260,9 +1260,22 @@ async fn pickup_one_ready_task(deps: &AgentDeps, logger: &CortexLogger) -> anyho
     let links = deps.links.clone();
     let agent_names = deps.agent_names.clone();
     let sqlite_pool = deps.sqlite_pool.clone();
+    let secrets_snapshot = deps.runtime_config.secrets.load().clone();
     tokio::spawn(async move {
+        // Helper closure: scrub both known secrets (Layer 1) and unknown leak
+        // patterns (Layer 2) from text before it reaches channels or events.
+        let scrub = |text: String| -> String {
+            let scrubbed = if let Some(store) = secrets_snapshot.as_ref() {
+                crate::secrets::scrub::scrub_with_store(&text, store)
+            } else {
+                text
+            };
+            crate::secrets::scrub::scrub_leaks(&scrubbed)
+        };
+
         match worker.run().await {
-            Ok(result_text) => {
+            Ok(raw_result) => {
+                let result_text = scrub(raw_result);
                 let db_updated = task_store
                     .update(
                         &agent_id,
@@ -1325,7 +1338,7 @@ async fn pickup_one_ready_task(deps: &AgentDeps, logger: &CortexLogger) -> anyho
             }
             Err(error) => {
                 let (error_message, notify, success) = map_worker_completion_result(Err(
-                    WorkerCompletionError::failed(error.to_string()),
+                    WorkerCompletionError::failed(scrub(error.to_string())),
                 ));
                 run_logger.log_worker_completed(worker_id, &error_message, false);
 
diff --git a/src/hooks/spacebot.rs b/src/hooks/spacebot.rs
index 934797d33..6ea610021 100644
--- a/src/hooks/spacebot.rs
+++ b/src/hooks/spacebot.rs
@@ -250,18 +250,54 @@ impl SpacebotHook {
     }
 
     /// Apply shared safety checks for tool output before any downstream handling.
+    ///
+    /// For channels, a detected secret terminates the agent immediately to prevent
+    /// exfiltration via the `reply` tool. For workers and branches, secrets are
+    /// logged but execution continues — these processes cannot communicate with
+    /// users directly, and their egress paths (worker results, branch conclusions,
+    /// status updates) apply scrubbing before content reaches the channel.
     pub(crate) fn guard_tool_result(&self, tool_name: &str, result: &str) -> HookAction {
         if let Some(leak) = self.scan_for_leaks(result) {
-            tracing::error!(
-                process_id = %self.process_id,
-                tool_name = %tool_name,
-                leak_prefix = %&leak[..leak.len().min(8)],
-                "secret leak detected in tool output, terminating agent"
-            );
-            return HookAction::Terminate {
-                reason: "Tool output contained a secret. Agent terminated to prevent exfiltration."
-                    .into(),
-            };
+            match self.process_type {
+                ProcessType::Worker | ProcessType::Branch => {
+                    // Workers and branches cannot communicate with users directly.
+                    // Their egress paths (worker results, branch conclusions,
+                    // status updates) scrub secrets before content reaches the
+                    // channel. Log and continue rather than killing the process.
+                    //
+                    // Avoid logging any fragment of the matched secret. Only log
+                    // the encoding kind (plaintext/url/base64/hex) and length.
+                    let kind = if leak.starts_with("url-encoded:") {
+                        "url-encoded"
+                    } else if leak.starts_with("base64-encoded:") {
+                        "base64-encoded"
+                    } else if leak.starts_with("hex-encoded:") {
+                        "hex-encoded"
+                    } else {
+                        "plaintext"
+                    };
+                    tracing::warn!(
+                        process_id = %self.process_id,
+                        tool_name = %tool_name,
+                        leak_kind = kind,
+                        leak_len = leak.len(),
+                        "secret detected in tool output (non-channel process, continuing)"
+                    );
+                }
+                ProcessType::Channel | ProcessType::Compactor | ProcessType::Cortex => {
+                    tracing::error!(
+                        process_id = %self.process_id,
+                        tool_name = %tool_name,
+                        leak_prefix = %&leak[..leak.len().min(8)],
+                        "secret leak detected in tool output, terminating agent"
+                    );
+                    return HookAction::Terminate {
+                        reason:
+                            "Tool output contained a secret. Agent terminated to prevent exfiltration."
+                                .into(),
+                    };
+                }
+            }
         }
 
         HookAction::Continue
@@ -488,9 +524,10 @@ where
             return guard_action;
         }
 
-        // Only enforce hard-stop leak blocking on channel egress (`reply`).
-        // Worker and branch tool outputs are internal and should not terminate
-        // long-running jobs.
+        // Belt-and-suspenders check specifically for `reply` tool results on
+        // channels. `guard_tool_result` already terminates channels on any tool
+        // leak, but this catches any edge case where the reply content itself
+        // has a different leak than the raw tool output.
         if self.process_type == ProcessType::Channel
             && tool_name == "reply"
             && let Some(leak) = self.scan_for_leaks(result)
@@ -508,8 +545,17 @@ where
         }
 
         // Cap the result stored in the broadcast event to avoid blowing up
-        // event subscribers with multi-MB tool results.
-        self.emit_tool_completed_event(tool_name, result);
+        // event subscribers with multi-MB tool results. For worker/branch
+        // processes, scrub leak patterns from the event payload so secrets
+        // don't reach the SSE dashboard.
+        if matches!(self.process_type, ProcessType::Worker | ProcessType::Branch) {
+            let scrubbed = crate::secrets::scrub::scrub_leaks(result);
+            let capped =
+                crate::tools::truncate_output(&scrubbed, crate::tools::MAX_TOOL_OUTPUT_BYTES);
+            self.emit_tool_completed_event_from_capped(tool_name, capped);
+        } else {
+            self.emit_tool_completed_event(tool_name, result);
+        }
 
         tracing::debug!(
             process_id = %self.process_id,
diff --git a/src/secrets/scrub.rs b/src/secrets/scrub.rs
index 0df22e4f5..3ffc354a3 100644
--- a/src/secrets/scrub.rs
+++ b/src/secrets/scrub.rs
@@ -33,6 +33,14 @@ static LEAK_PATTERNS: LazyLock<Vec<Regex>> = LazyLock::new(|| {
     ]
 });
 
+/// Full PEM block pattern (header + base64 body + footer). Used by `scrub_leaks()`
+/// to redact entire private key blocks, not just the header line matched by
+/// `LEAK_PATTERNS`.
+static PEM_BLOCK: LazyLock<Regex> = LazyLock::new(|| {
+    Regex::new(r"(?s)-----BEGIN[A-Z \r\n]*PRIVATE KEY-----.*?-----END[A-Z \r\n]*PRIVATE KEY-----")
+        .expect("hardcoded regex")
+});
+
 static BASE64_SEGMENT: LazyLock<Regex> =
     LazyLock::new(|| Regex::new(r"[A-Za-z0-9+/]{24,}={0,2}").expect("hardcoded regex"));
 
@@ -222,6 +230,30 @@ pub fn scrub_with_store(text: &str, store: &crate::secrets::store::SecretsStore)
     scrub_secrets(text, &pairs)
 }
 
+/// Replace all detected leak patterns in `content` with a redaction marker.
+///
+/// Unlike `scan_for_leaks()` which returns the first match, this function
+/// replaces **all** plaintext matches of known API key patterns with
+/// `[LEAKED_SECRET_REDACTED]`. Used on egress paths (worker results, branch
+/// conclusions, status updates) to sanitize content that may contain secrets
+/// the LLM encountered during tool execution.
+///
+/// Does NOT check encoded forms (base64, hex, URL-encoded) — those are
+/// unlikely to appear in LLM-generated output text.
+pub fn scrub_leaks(content: &str) -> String {
+    // First, redact full PEM blocks (header + body + footer) so the base64
+    // key material is removed, not just the header line.
+    let mut result = PEM_BLOCK
+        .replace_all(content, "[LEAKED_SECRET_REDACTED]")
+        .into_owned();
+    for pattern in LEAK_PATTERNS.iter() {
+        result = pattern
+            .replace_all(&result, "[LEAKED_SECRET_REDACTED]")
+            .into_owned();
+    }
+    result
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -284,4 +316,68 @@ mod tests {
         let result = scan_for_leaks("this is normal output with no secrets");
         assert!(result.is_none());
     }
+
+    #[test]
+    fn scrub_leaks_redacts_anthropic_key() {
+        let input = "found key sk-ant-abc123456789012345678 in config";
+        let result = scrub_leaks(input);
+        assert!(
+            !result.contains("sk-ant-abc123456789012345678"),
+            "secret should be redacted in: {result}"
+        );
+        assert!(
+            result.contains("[LEAKED_SECRET_REDACTED]"),
+            "redaction marker missing in: {result}"
+        );
+        assert!(
+            result.contains("found key"),
+            "surrounding text should be preserved in: {result}"
+        );
+    }
+
+    #[test]
+    fn scrub_leaks_redacts_multiple_secrets() {
+        let input =
+            "keys: sk-ant-abc123456789012345678 and ghp_abcdefghijklmnopqrstuvwxyz0123456789";
+        let result = scrub_leaks(input);
+        assert!(
+            !result.contains("sk-ant-"),
+            "first secret should be redacted in: {result}"
+        );
+        assert!(
+            !result.contains("ghp_"),
+            "second secret should be redacted in: {result}"
+        );
+    }
+
+    #[test]
+    fn scrub_leaks_passes_through_clean_text() {
+        let input = "this is normal output with no secrets";
+        assert_eq!(scrub_leaks(input), input);
+    }
+
+    #[test]
+    fn scrub_leaks_redacts_full_pem_block() {
+        let input = "config:\n-----BEGIN RSA PRIVATE KEY-----\n\
+                      MIIEpAIBAAKCAQEA0Z3VS5JJcds3xfn/ygWyF8PbnGy0AHB7MhgHcTz6sE2I2yPB\n\
+                      aFDrBz9vFqU4zK3L3hUfVnEy\n\
+                      -----END RSA PRIVATE KEY-----\nmore text";
+        let result = scrub_leaks(input);
+        assert!(
+            !result.contains("MIIEpAIBAAK"),
+            "PEM body should be redacted in: {result}"
+        );
+        assert!(
+            !result.contains("BEGIN RSA PRIVATE KEY"),
+            "PEM header should be redacted in: {result}"
+        );
+        assert!(
+            result.contains("[LEAKED_SECRET_REDACTED]"),
+            "redaction marker missing in: {result}"
+        );
+        assert!(
+            result.contains("more text"),
+            "surrounding text should be preserved in: {result}"
+        );
+    }
 }
diff --git a/src/tools/set_status.rs b/src/tools/set_status.rs
index 2b50ed9c6..23633de2e 100644
--- a/src/tools/set_status.rs
+++ b/src/tools/set_status.rs
@@ -101,7 +101,10 @@ impl Tool for SetStatusTool {
         };
 
         // Scrub tool secret values before the status reaches the channel.
+        // Layer 1: exact-match redaction of known secrets from the store.
+        // Layer 2: regex-based redaction of unknown secret patterns.
         let status = crate::secrets::scrub::scrub_secrets(&status, &self.tool_secret_pairs);
+        let status = crate::secrets::scrub::scrub_leaks(&status);
 
         let event = ProcessEvent::WorkerStatus {
             agent_id: self.agent_id.clone(),