Claude Code Engineer Thariq Admits to Implanting Detection Backdoors and Spy Code Targeting Users (Especially Chinese Users) in March Update
Thariq stated that this move aims to prevent model abuse and distillation, and the related code will be rolled back tomorrow to address the issue.
This incident exposes Anthropic's challenge in balancing security measures and user privacy.
In market mechanisms, developers and users assess the trustworthiness of Claude Code as buyers, and the incident prompted Thariq to publicly acknowledge it, leading to funding flowing towards competing AI code tools; Anthropic faces short-term reputation risks, with security-focused users shifting to more transparent platforms.
Source: Public Information
ABAB AI Insight
Anthropic has previously strengthened security classifiers and anti-distillation measures in the Claude series models, similar to OpenAI's evolution of jailbreak protections for GPT models. This backdoor incident reflects the defensive deployments of leading AI companies under global regulatory and technological diffusion pressures.
In terms of capital pathways, Anthropic is enhancing model protection through internal code control and government collaboration to mobilize resources, motivated by the need to maintain intellectual property and national security compliance while avoiding large-scale distillation that could leak capabilities and impact commercial valuations.
Similar to Google's early anti-scraping and data protection mechanisms or Meta's restrictions on the open-sourcing of the Llama model, the AI code tools industry is in a transitional phase of balancing security backdoors and user trust.
Essentially, this is a response to regulatory changes: AI developers are adopting covert detection mechanisms to prevent abuse and distillation pressures, with mechanisms tailored to regional differentiated protections against geopolitical technological competition, driving the industry from open innovation towards a compliance-first code governance framework and reshaping global AI development standards.
ABAB News · Cognitive Laws
Backdoor anti-distillation, trust is the first to erode.
In a security-first environment, user choices are votes.
Rolling back code is easy, but repairing reputation is difficult.