Unverified50% confidenceEventExact time
In April 2025, researchers at Invariant Labs demonstrated that tool descriptions can carry instructions and be executed by AI; in the test, an attacker-controlled server, through hidden instructions, sent chat records to the attacker's number at the same time the user sent a WhatsApp message.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAttackers tamper with a model's chat template by calling the create API, permanently injecting malicious instructions into the System prompt; this change is stored in the model's Modelfile and is invisible to the client and undetectable by standard tools.70% similarUnverifiedThe SentryFlow demo client has two sessions, runs on protocol version 2025-11-25, and exposes 17 tools via tools/list.66% similarUnverifiedClosed-source labs can intervene in model behavior after the fact through access controls, terms of use, and API rate limiting.65% similarVerifiedAgent mode is implemented through a 'tool-calling' mechanism, whereby the model can autonomously decide to call external APIs, execute shell commands, read and write files, or browse the web.64% similarUnverifiedAgentic coding tools can autonomously read and write files, execute terminal commands, break down tasks and advance them step by step, and require frequent requests for human confirmation to control risk64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1028970API
curl https://kongchang.com/api/v1/knowledge/claims/1028970MCP
get_claim(id=1028970)