Open-source AI browser agent for Chrome and Firefox (monorepo) 🧠
JavaScript ai-agentai-agentsai-toolsbrowser-aibrowser-automationbrowser-extensionchrome-aichrome-devtoolschrome-extensionchrome-extensionsfirefox-addonfirefox-addonsfirefox-extensionfirefox-extensionsgemmagemma4minimaxoffline-aioffline-llmqwen
fix(guard): validate recipient semantics on displayed question regardless of declared purpose
Emre Sokullu7a50628a5941
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.jsindex bf5cd18f8..372a7fd06 100644--- a/src/chrome/src/agent/agent.js+++ b/src/chrome/src/agent/agent.js@@ -17899,20 +17899,18 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d_isRecipientClarification(context, pendingType = 'message_recipient', observedCandidates = []) {if (!context || typeof context !== 'object') return false;const purpose = String(context.purpose || '').trim();+ if (purpose === 'research_escalation') return false;++ const questionText = [+ context.question,+ ...(Array.isArray(context.options) ? context.options : []),+ ].filter(Boolean).join(' ').toLowerCase();+if (pendingType === 'recipient_change') {- if (purpose === 'recipient_change') return true;- if (purpose === 'research_escalation') return false;- const text = [- context.question,- context.reason,- ...(Array.isArray(context.options) ? context.options : []),- ].filter(Boolean).join(' ').toLowerCase();- const changeWord = /(change|switch|different|instead|update|replace|another|new|更改|更换|切换|替换|换成)/i.test(text);- const targetWord = /(recipient|conversation|contact|send to|person|address|target|someone else|\b(?:to|cc|bcc|role|field)\b|收件人|联系人|发送)/i.test(text);+ const changeWord = /(change|switch|different|instead|update|replace|another|new|更改|更换|切换|替换|换成)/i.test(questionText);+ const targetWord = /(recipient|conversation|contact|send to|person|address|target|someone else|\b(?:to|cc|bcc|role|field)\b|收件人|联系人|发送)/i.test(questionText);return changeWord && targetWord;}- if (purpose === 'message_recipient' || purpose === 'recipient_change') return true;- if (purpose === 'research_escalation') return false;// Check if the clarify question, reason, or options explicitly mention any observed candidate identity or aliasif (Array.isArray(observedCandidates) && observedCandidates.length > 0) {@@ -17933,16 +17931,10 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d}}- const text = [- context.question,- context.reason,- ...(Array.isArray(context.options) ? context.options : []),- ].filter(Boolean).join(' ').toLowerCase();-- if (/recipient|收件人/i.test(text)) return true;+ if (/recipient|收件人/i.test(questionText)) return true;- const hasRecipientTerm = /(contact|person|address|who|whom|which contact|which person|which recipient|which user|which address|to whom|for whom|someone|联系人|哪位|谁)/i.test(text);- const hasMessagingTerm = /(send|receive|deliver|message|email|mail|draft|reply|chat|conversation|发送|邮件|信息|消息|寄|回复)/i.test(text);+ const hasRecipientTerm = /(contact|person|address|who|whom|which contact|which person|which recipient|which user|which address|to whom|for whom|someone|联系人|哪位|谁)/i.test(questionText);+ const hasMessagingTerm = /(send|receive|deliver|message|email|mail|draft|reply|chat|conversation|发送|邮件|信息|消息|寄|回复)/i.test(questionText);return hasRecipientTerm && hasMessagingTerm;}diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.jsindex 5a480f418..8d9a0c9a8 100644--- a/src/firefox/src/agent/agent.js+++ b/src/firefox/src/agent/agent.js@@ -15652,20 +15652,18 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d_isRecipientClarification(context, pendingType = 'message_recipient', observedCandidates = []) {if (!context || typeof context !== 'object') return false;const purpose = String(context.purpose || '').trim();+ if (purpose === 'research_escalation') return false;++ const questionText = [+ context.question,+ ...(Array.isArray(context.options) ? context.options : []),+ ].filter(Boolean).join(' ').toLowerCase();+if (pendingType === 'recipient_change') {- if (purpose === 'recipient_change') return true;- if (purpose === 'research_escalation') return false;- const text = [- context.question,- context.reason,- ...(Array.isArray(context.options) ? context.options : []),- ].filter(Boolean).join(' ').toLowerCase();- const changeWord = /(change|switch|different|instead|update|replace|another|new|更改|更换|切换|替换|换成)/i.test(text);- const targetWord = /(recipient|conversation|contact|send to|person|address|target|someone else|\b(?:to|cc|bcc|role|field)\b|收件人|联系人|发送)/i.test(text);+ const changeWord = /(change|switch|different|instead|update|replace|another|new|更改|更换|切换|替换|换成)/i.test(questionText);+ const targetWord = /(recipient|conversation|contact|send to|person|address|target|someone else|\b(?:to|cc|bcc|role|field)\b|收件人|联系人|发送)/i.test(questionText);return changeWord && targetWord;}- if (purpose === 'message_recipient' || purpose === 'recipient_change') return true;- if (purpose === 'research_escalation') return false;// Check if the clarify question, reason, or options explicitly mention any observed candidate identity or aliasif (Array.isArray(observedCandidates) && observedCandidates.length > 0) {@@ -15686,16 +15684,10 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d}}- const text = [- context.question,- context.reason,- ...(Array.isArray(context.options) ? context.options : []),- ].filter(Boolean).join(' ').toLowerCase();-- if (/recipient|收件人/i.test(text)) return true;+ if (/recipient|收件人/i.test(questionText)) return true;- const hasRecipientTerm = /(contact|person|address|who|whom|which contact|which person|which recipient|which user|which address|to whom|for whom|someone|联系人|哪位|谁)/i.test(text);- const hasMessagingTerm = /(send|receive|deliver|message|email|mail|draft|reply|chat|conversation|发送|邮件|信息|消息|寄|回复)/i.test(text);+ const hasRecipientTerm = /(contact|person|address|who|whom|which contact|which person|which recipient|which user|which address|to whom|for whom|someone|联系人|哪位|谁)/i.test(questionText);+ const hasMessagingTerm = /(send|receive|deliver|message|email|mail|draft|reply|chat|conversation|发送|邮件|信息|消息|寄|回复)/i.test(questionText);return hasRecipientTerm && hasMessagingTerm;}diff --git a/test/run.js b/test/run.jsindex c0a70e75f..3c1431ea2 100644--- a/test/run.js+++ b/test/run.js@@ -5772,6 +5772,29 @@ test('direct-message recipient guard uses structured intent and exact active iderecipients: [{ identity: '[email protected]', role: 'to' }],}, `${label}: recipient change did not update target`);+ // A mislabeled question with purpose: 'recipient_change' but lacking recipient-change semantics+ // must never authorize changing the recipient.+ agent._planExecutionGuards.set(tabId, recipientChangeGuardState([{ identity: '[email protected]', role: 'to' }]));+ assert.equal(+ agent._bindClarifiedMessageRecipient(tabId, '[email protected]', 'user', {+ question: 'Which colleague should present?',+ purpose: 'recipient_change',+ }),+ false,+ `${label}: mislabeled clarification question with purpose: recipient_change authorized recipient change`,+ );++ // A mislabeled question with purpose: 'message_recipient' must never authorize a standard recipient either.+ agent._planExecutionGuards.set(tabId, clarifyGuardState([{ identity: '[email protected]', role: 'to' }]));+ assert.equal(+ agent._bindClarifiedMessageRecipient(tabId, '[email protected]', 'user', {+ question: 'Which colleague should present?',+ purpose: 'message_recipient',+ }),+ false,+ `${label}: mislabeled clarification question with purpose: message_recipient authorized recipient`,+ );+// Role preservation during recipient-change clarification:// A plan authorizing Alice as BCC when composer exposes Alice as To must retain BCC// on an identity-only answer, preventing unauthorized exposure on next send.