Expired50% confidenceEventExact time
On January 14, 2026, PromptArmor published an experiment showing that an attacker, by uploading a Word file containing hidden instructions, exploited Anthropic's own API in the VM whitelist to make a curl call a legitimate outbound request and exfiltrate private files from the local machine, with no manual approval required during the process.
1
Sources
50%
Confidence
Short-term (~14 days)
Relevance
10/15/2026
First Seen
This claim has expired. It is kept for the record and no longer describes the current state.
Valid until: 10/29/2026(expired)
Sources
Related Entities
Related Claims
UnverifiedBy 2024, security researchers had demonstrated indirect prompt injection attacks that hijack RAG system outputs by embedding hidden instructions in web pages70% similarUnverifiedIn 2023, researchers demonstrated hiding malicious instructions in invisible text areas of web pages, successfully hijacking a GPT-4-based browser agent into performing unauthorized operations.65% similarUnverifiedInvariant Labs在2025年4月1日首次将描述中夹带客户端界面不可见指令的现象命名为'工具投毒'（tool poisoning）。64% similarVerifiedIndirect injection refers to attackers hiding malicious instructions in external content that the model will read, such as webpage body text, email body text, or document content.64% similarUnverifiedAttackers can hide malicious code in postinstall scripts, causing it to execute in the user's environment when downstream projects install dependencies, while remaining transparent to approval processes that only check package names and version numbers64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1033347API
curl https://kongchang.com/api/v1/knowledge/claims/1033347MCP
get_claim(id=1033347)