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(agent): exclude authored links from alias proof and localize bare destinations
Emre Sokullu56ffacd836f7
diff --git a/src/chrome/src/agent/agent.js b/src/chrome/src/agent/agent.jsindex 792f726da..373d82bcd 100644--- a/src/chrome/src/agent/agent.js+++ b/src/chrome/src/agent/agent.js@@ -2448,7 +2448,11 @@ export class Agent extends LoopDetector {if (!intended || !published || intended === published) return false;const isDid = value => /^bluesky:did:/i.test(String(value || ''));if (isDid(intended) === isDid(published)) return false;+ // Links the author wrote in the post body are content, not proof of who+ // wrote the post. A wrong-account post that literally links to the+ // requested DID would otherwise certify itself as an alias.const cardAccounts = new Set((Array.isArray(record?.links) ? record.links : [])+ .filter(link => !link?.authored).flatMap(link => [link?.href, link?.expandedUrl]).map(value => this._workflowSocialPublicationAccountIdentity(siteWorkflow, value)).filter(value => !!value && isDid(value) !== isDid(published)));@@ -14624,14 +14628,26 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d// posts on X, then share the findings in the survey" mentions both a// publish verb and X without ever asking for a post. The verb has to// govern the platform through a destination preposition for this to be a- // publication target.+ // publication target. Verbs and prepositions are the same multilingual+ // sets the URL scan uses, so "Publícalo en X" counts while English-only+ // matching would leave it unbound.const publishesTo = (platform) => {- const pattern = new RegExp(- `\\b(?:post|publish|share|tweet|skeet)\\b[^.!?\\n]{0,60}?\\b(?:to|on|onto|in|at|via)\\s+(?:the\\s+)?${platform}\\b`,- 'i',+ const platformPattern = new RegExp(+ `(?<![${SOCIAL_WORD_EDGE}])(?:on|onto|to|via|in|at|en|sur|sobre|\u00e0|au|auf|su|em|na|para|\u0432|\u043d\u0430|\u5230)(?![${SOCIAL_WORD_EDGE}])\\s+(?:the\\s+)?${platform}(?![${SOCIAL_WORD_EDGE}])`,+ 'iu',);- return trustedContext.split(SOCIAL_CLAUSE_BREAK)- .some(clause => !!clause && pattern.test(clause) && this._socialPublicationCommandIn(clause));+ const verbPattern = new RegExp(SOCIAL_PUBLISH_VERBS.source, 'giu');+ return trustedContext.split(SOCIAL_CLAUSE_BREAK).some((clause) => {+ if (!clause) return false;+ for (const verb of clause.matchAll(verbPattern)) {+ const verbIndex = verb.index ?? 0;+ if (SOCIAL_READ_VERBS.test(clause.slice(0, verbIndex))) continue;+ const afterVerb = clause.slice(verbIndex, verbIndex + (verb[0]?.length ?? 0) + 60);+ platformPattern.lastIndex = 0;+ if (platformPattern.test(afterVerb)) return true;+ }+ return false;+ });};if (publishesTo('(?:bluesky|bsky(?:\\.app)?)')) targets.add('bluesky');if (publishesTo('(?:x|x\\.com|twitter(?:\\.com)?)')@@ -27931,7 +27947,11 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d].filter(candidate => !isEmbedded(candidate)).slice(0, 100).map(candidate => {let href = '';try { href = new URL(candidate.getAttribute('href') || candidate.href || '', location.href).href; } catch {}- const compact = value => String(value || '').replace(/\\s+/g, ' ').trim().slice(0, 1000);+ const compact = value => String(value || '').replace(/\s+/g, ' ').trim().slice(0, 1000);+ let inAuthoredBody = false;+ try {+ inAuthoredBody = authoredNodes.some(node => node === candidate || node.contains?.(candidate));+ } catch {}return {href: compact(href),text: compact(candidate.innerText || candidate.textContent),@@ -27942,10 +27962,11 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d|| candidate.getAttribute('data-full-url')|| candidate.getAttribute('data-url'),),+ ...(inAuthoredBody ? { authored: true } : {}),};}).filter(candidate => candidate.href || candidate.text || candidate.expandedUrl);const links = Array.from(new Map(recordLinks.map(candidate => [- [candidate.href, candidate.text, candidate.title, candidate.ariaLabel, candidate.expandedUrl].join('\\u0000'),+ [candidate.href, candidate.text, candidate.title, candidate.ariaLabel, candidate.expandedUrl, candidate.authored ? '1' : ''].join('\u0000'),candidate,])).values()).slice(0, 24);const prior = workflowResourceRecordMap.get(identity);diff --git a/src/firefox/src/agent/agent.js b/src/firefox/src/agent/agent.jsindex 1d839197e..273b5b60b 100644--- a/src/firefox/src/agent/agent.js+++ b/src/firefox/src/agent/agent.js@@ -2162,7 +2162,11 @@ export class Agent extends LoopDetector {if (!intended || !published || intended === published) return false;const isDid = value => /^bluesky:did:/i.test(String(value || ''));if (isDid(intended) === isDid(published)) return false;+ // Links the author wrote in the post body are content, not proof of who+ // wrote the post. A wrong-account post that literally links to the+ // requested DID would otherwise certify itself as an alias.const cardAccounts = new Set((Array.isArray(record?.links) ? record.links : [])+ .filter(link => !link?.authored).flatMap(link => [link?.href, link?.expandedUrl]).map(value => this._workflowSocialPublicationAccountIdentity(siteWorkflow, value)).filter(value => !!value && isDid(value) !== isDid(published)));@@ -12424,14 +12428,26 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d// posts on X, then share the findings in the survey" mentions both a// publish verb and X without ever asking for a post. The verb has to// govern the platform through a destination preposition for this to be a- // publication target.+ // publication target. Verbs and prepositions are the same multilingual+ // sets the URL scan uses, so "Publícalo en X" counts while English-only+ // matching would leave it unbound.const publishesTo = (platform) => {- const pattern = new RegExp(- `\\b(?:post|publish|share|tweet|skeet)\\b[^.!?\\n]{0,60}?\\b(?:to|on|onto|in|at|via)\\s+(?:the\\s+)?${platform}\\b`,- 'i',+ const platformPattern = new RegExp(+ `(?<![${SOCIAL_WORD_EDGE}])(?:on|onto|to|via|in|at|en|sur|sobre|\u00e0|au|auf|su|em|na|para|\u0432|\u043d\u0430|\u5230)(?![${SOCIAL_WORD_EDGE}])\\s+(?:the\\s+)?${platform}(?![${SOCIAL_WORD_EDGE}])`,+ 'iu',);- return trustedContext.split(SOCIAL_CLAUSE_BREAK)- .some(clause => !!clause && pattern.test(clause) && this._socialPublicationCommandIn(clause));+ const verbPattern = new RegExp(SOCIAL_PUBLISH_VERBS.source, 'giu');+ return trustedContext.split(SOCIAL_CLAUSE_BREAK).some((clause) => {+ if (!clause) return false;+ for (const verb of clause.matchAll(verbPattern)) {+ const verbIndex = verb.index ?? 0;+ if (SOCIAL_READ_VERBS.test(clause.slice(0, verbIndex))) continue;+ const afterVerb = clause.slice(verbIndex, verbIndex + (verb[0]?.length ?? 0) + 60);+ platformPattern.lastIndex = 0;+ if (platformPattern.test(afterVerb)) return true;+ }+ return false;+ });};if (publishesTo('(?:bluesky|bsky(?:\\.app)?)')) targets.add('bluesky');if (publishesTo('(?:x|x\\.com|twitter(?:\\.com)?)')@@ -24405,7 +24421,11 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d].filter(candidate => !isEmbedded(candidate)).slice(0, 100).map(candidate => {let href = '';try { href = new URL(candidate.getAttribute('href') || candidate.href || '', location.href).href; } catch {}- const compact = value => String(value || '').replace(/\\s+/g, ' ').trim().slice(0, 1000);+ const compact = value => String(value || '').replace(/\s+/g, ' ').trim().slice(0, 1000);+ let inAuthoredBody = false;+ try {+ inAuthoredBody = authoredNodes.some(node => node === candidate || node.contains?.(candidate));+ } catch {}return {href: compact(href),text: compact(candidate.innerText || candidate.textContent),@@ -24416,10 +24436,11 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d|| candidate.getAttribute('data-full-url')|| candidate.getAttribute('data-url'),),+ ...(inAuthoredBody ? { authored: true } : {}),};}).filter(candidate => candidate.href || candidate.text || candidate.expandedUrl);const links = Array.from(new Map(recordLinks.map(candidate => [- [candidate.href, candidate.text, candidate.title, candidate.ariaLabel, candidate.expandedUrl].join('\\u0000'),+ [candidate.href, candidate.text, candidate.title, candidate.ariaLabel, candidate.expandedUrl, candidate.authored ? '1' : ''].join('\u0000'),candidate,])).values()).slice(0, 24);const prior = workflowResourceRecordMap.get(identity);diff --git a/test/run.js b/test/run.jsindex a85bdf503..e15e68e81 100644--- a/test/run.js+++ b/test/run.js@@ -85044,6 +85044,13 @@ test('social publication workflow follows the live X or Bluesky destination and'a read verb further from the noun did not disqualify it'],['Read the summary and publish it on https://x.com/home', ['twitter'],'a publish command in its own clause was lost to an earlier read verb'],+ // Bare platforms use the same multilingual verbs as URLs, not English-only.+ ['Publícalo en X', ['twitter'],+ 'a Spanish bare platform destination was missed'],+ ['Publicar en Bluesky', ['bluesky'],+ 'a Spanish bare Bluesky destination was missed'],+ ['Publiez ceci sur Bluesky', ['bluesky'],+ 'a French bare platform destination was missed'],]) {assert.deepEqual([...agent._trustedSocialPublishTargetAdapters({ taskText })],@@ -85448,6 +85455,31 @@ test('a Bluesky DID and handle name one account only on the card that proves it'handle,cardWithDid,), false, AgentClass.name + ': a different handle was accepted as an alias');++ // A link the author wrote in the post body is content, not proof of who+ // wrote the post. A wrong-account post linking to the requested DID must+ // not certify itself.+ assert.equal(agent._workflowSocialAccountAliasProven(+ workflow,+ did,+ 'bluesky:attacker.one',+ {+ url: 'https://bsky.app/profile/attacker.one/post/9xyz',+ links: [{ href: 'https://bsky.app/profile/did:plc:abc123', authored: true }],+ },+ ), false, AgentClass.name + ': a body link to the intended DID was accepted as alias proof');+ assert.equal(agent._workflowSocialAccountAliasProven(+ workflow,+ did,+ 'bluesky:attacker.one',+ {+ url: 'https://bsky.app/profile/attacker.one/post/9xyz',+ links: [+ { href: 'https://bsky.app/profile/attacker.one' },+ { href: 'https://bsky.app/profile/did:plc:abc123', authored: true },+ ],+ },+ ), false, AgentClass.name + ': an authored DID link bridged two unrelated accounts');assert.equal(agent._workflowSocialAccountAliasProven(agent._resolvePlannerSiteWorkflow('https://x.com/home', {