According to a report by ithome.com.tw dated August 7, 2026, security firm Malwarebytes said Anthropic's AI chatbot Claude has a share-chat feature that can result in conversations being found through Google Search, exposing sensitive information contained in those chats (E1). The report traces the root cause to Claude's Share Chats function: citing Anthropic's own documentation, the feature generates a shareable link for a conversation and simultaneously changes that conversation's status to "Public" (E3). Once a chat is set to Public via a share link, it becomes accessible to anyone who has the link — and, as the report describes, also becomes reachable by search engine crawlers.
The ithome.com.tw report states that the issue was first identified by Reddit users, who found that using specific search techniques on Google turned up numerous Claude conversation logs in search results (E2). This user-driven discovery — rather than a formal security audit — was how the exposure came to public attention, according to the report.
Malwarebytes told ithome.com.tw that this is not a Claude-only problem. The firm said it had previously found conversation logs from xAI's Grok appearing in Google search results, and that Meta AI had exhibited a similar issue in the past (E4). This places the Claude share-link exposure within a pattern the security firm says it has observed across more than one AI chat product, based on its own prior findings.
Per the same report, Malwarebytes' recommendation is direct: the simplest way to avoid sensitive data leaking through this kind of shared-conversation exposure is to not share AI conversations with anyone at all, or to avoid entering personally identifiable information into an AI chat in the first place (E5). The firm frames this as a user-behavior mitigation rather than something that depends on a platform fix.
A separate report on futurecity.cw.com.tw, dated May 21, 2026, describes how default "zero training" configurations vary by platform. It notes that ChatGPT's default setting has training turned on, requiring users to manually switch it off; Claude similarly requires users to confirm and adjust the setting; and for Gemini, the personal version offers no way to adjust this setting at all, while the enterprise version defaults to it being off (E9). This means the baseline risk of a user's input being retained or used for training differs depending on which product and which tier (personal vs. enterprise) a user is on, according to the report.
The same futurecity.cw.com.tw report cites the example of OpenClaw, where it says many users mistakenly believe that model computation is happening locally on their Mac Mini device. In reality, according to the report, information is still sent out to an external AI model — and it notes that some users have directly entered highly sensitive personal data, including health insurance card information, into the tool under this false assumption (E10).
Taken together, the evidence points to two distinct but related exposure paths. First, Claude's share-chat design (E1, E3) turns a convenience feature — generating a link to show someone else a conversation — into a public, search-indexable page, a pattern Malwarebytes says it has also seen with Grok and Meta AI (E4), and which was first surfaced not by a formal audit but by Reddit users searching Google (E2). Malwarebytes' own recommended fix for this — simply not sharing chats or personal data (E5) — addresses the sharing-link vector specifically. Second, separate reporting on default "zero training" settings (E9) and on user misunderstanding about where computation actually occurs, as in the OpenClaw case (E10), suggests that avoiding the share button alone would not close every exposure gap: even conversations that are never shared via a link can still leave the device and reach an external model, depending on platform defaults and user assumptions about local versus remote processing.
No. According to the ithome.com.tw report, Malwarebytes said it had previously found Grok conversation logs appearing in Google search results, and that Meta AI had a similar issue in the past.
Per the ithome.com.tw report, the issue was first identified by Reddit users, who used specific search techniques on Google to locate multiple Claude conversation logs.
Malwarebytes told ithome.com.tw that the simplest approach is to not share AI conversations with anyone, or to avoid entering personally identifiable information into AI chats at all.
According to TechNews and CNA, Winbond Electronics (華邦電) GM Chen Pei-ming warned on August 6, 2026 that 2027 memory shortages will exceed 2026's. CMS revenue rose 78% and Flash revenue rose 65% quarter-over-quarter, backed by NT$39.5 billion in 2026 capex, up over 70-fold year-on-year.
According to Technews, MoneyDJ, and CTEE reports dated August 6, 2026, TSMC (台積電) is holding roughly $1 billion (about NT$32 billion) worth of Apple (蘋果) A20 Pro processors that cannot complete back-end packaging because of a DRAM shortage. TSMC's CFO had already flagged rising inventory days on July 16, and Apple sources memory mainly from Micron, plus SK Hynix and Samsung.
According to Technews, also reported by CNA and ETtoday, the US FCC is drafting a measure to ban imports of new Chinese-made optical transceivers, with officials hoping to finalize it this year, though it could still be revised or shelved. The news drove same-day share drops of 7.27% for Innolight (中際旭創) and 5.29% for Eoptolink (新易盛), while Beijing's foreign ministry called the move protectionist.