Why is Claude Code considered 'dumber' or less effective now?

Claude Code's performance significantly degraded since early March, particularly after version 2.1.69. It now exhibits increased 'laziness' (stop-hook violations), reduced engagement with code, and a tendency to rewrite entire files instead of making targeted edits. This makes it unreliable for complex engineering tasks.

Who reported the performance issues with Anthropic's Claude Code?

Stella Laurenzo, the director of the AI group at AMD, initiated the complaint on GitHub and LinkedIn. Her team's analysis of thousands of Claude Code sessions confirmed a steep decline in reasoning quality and reliability.

AMD AI Director Blasts 'Dumber' Claude Code: Performance & Security Flaws

Q: What specific feature change contributed to Claude Code's degradation?

The degradation is attributed to the early March deployment of Claude Code version 2.1.69, which introduced 'thinking content redaction.' This feature strips the AI's internal thought process from API responses, making its reasoning opaque to users.

Q: Are there any security vulnerabilities associated with Claude Code?

Yes, following a source code leak, Adversa AI discovered a critical vulnerability. This flaw allows malicious commands to bypass security checks within a 50+ subcommand pipeline, potentially leading to unauthorized data exfiltration.

Anthropic's Claude Code, once a favored AI assistant for complex engineering tasks, now faces accusations of significant performance degradation. An official GitHub issue opened by AMD's AI group director details a steep decline in Claude Code's reasoning quality and reliability since early March, prompting the team to switch providers, according to The Register . This erosion of trust raises critical questions about the stability and future trajectory of advanced AI coding tools.

What Triggered the Performance Drop?

Stella Laurenzo, director of the AI group at chipmaker AMD, initiated the discussion with a detailed complaint on GitHub, supported by a LinkedIn post . Her team concluded that Claude Code "cannot be trusted to perform complex engineering tasks" after months of consistent use in a high-complexity work environment. Laurenzo's assessment, echoed by other senior engineers on her team, points to a clear drop-off in the AI's efficacy.

Their analysis of 6,852 Claude Code sessions, which included 234,760 tool calls and 17,871 thinking blocks, revealed alarming trends. The number of "stop-hook violations"—indicators of the AI's "laziness," such as prematurely stopping thinking processes or avoiding responsibility—skyrocketed. This metric went from zero before March 8th to an average of 10 violations per day by the end of last month. Claude Code's engagement with code also decreased dramatically. The average number of times it would read a piece of code before making changes fell from 6.6 reads to just 2 by late March. Simultaneously, the AI began rewriting entire files more frequently instead of making targeted edits. Laurenzo attributes these changes directly to the early March deployment of Claude Code version 2.1.69, which introduced "thinking content redaction." This feature, enabled by default, strips the AI's internal thought process from API responses, leaving users unaware of Claude Code's reasoning.

Broader Issues and Security Concerns

The performance degradation isn't an isolated incident for Anthropic. Users previously reported issues in February when version 2.1.20 truncated explanations of its reading process. This earlier change prompted concerns that the AI was being "dumbed down." Beyond performance, Anthropic also faced criticism for unexplained surges in token usage, pushing some users past their limits and restricting product access. Adding to the company's challenges, Claude Code's source code was recently exposed. Following this leak, Adversa AI discovered a critical vulnerability. This flaw allows malicious commands to bypass security checks if presented within a 50+ subcommand pipeline, potentially leading to unauthorized data exfiltration. Melissa Bischoping, senior director of security & product design research at Tanium , noted that such leaks could allow adversaries to build lookalikes that install malware or harvest credentials, according to SecurityWeek. Laurenzo urged Anthropic for transparency regarding thinking token usage and requested a "max thinking tier" for complex workflows, asserting that engineers would pay more for guaranteed deep reasoning. While her team has transitioned to another provider, Laurenzo warned Anthropic that its declining performance risks surrendering its previous market leadership. Six months ago, Claude "stood alone in terms of reasoning quality," but competitors now challenge its position at the top.