Ben Horowitz calls Anthropic 'one of the most negligent companies' on security

Andreessen Horowitz co-founder Ben Horowitz called Anthropic "one of the most negligent companies that we've ever seen in terms of security" on the All-In podcast published on 17 October 2026 [1]. He backed the claim with an anecdote about an Opus 4.5 agent hacking a test website, which has not been independently verified [1][2].

What did Horowitz say about Anthropic?

Ben Horowitz called Anthropic one of the most negligent companies on security on the All-In podcast of 17 October 2026. He was filling in as a guest host, and All-In posted the passage as a clip on X later that day [1][2].

According to Horowitz, his firm ran cybersecurity checks when it first got Claude Opus 4.5 [1]. One agent built a website for a pet store, and a second agent was told to look for exploits in it [1].

Horowitz says an Opus 4.5 agent broke into a test pet-shop site and Anthropic told his firm it was prompted wrong. In his telling, the agent found a SQL injection flaw, answered "let's pop it open" when asked if it was exploitable, and took all the customer data [1].

Horowitz gave no date, test setup, prompt or record of the call [1]. The site was built by his firm's own agent for the test, and he did not say whether any real customer data was involved [1].

Has Anthropic responded?

Anthropic had not responded publicly as of 19 October 2026, 03:00 CEST, based on The AI Pin's check of the web and of posts on X by Anthropic, Claude, Dario Amodei and Jack Clark since 17 October. Other outlets covering the clip also reported no response [8].

On 9 October, before the episode, Anthropic published a report on unintended model actions that lists cases of Claude "exploiting a basic flaw in software to run commands on a server" during evaluations and internal use [7]. That report is related context, not a reply to Horowitz [7].

Why does the a16z connection matter?

Anthropic does not appear on a16z's published investment list, which includes OpenAI, xAI and Mistral. The list also names Safe Superintelligence, Thinking Machines Lab, Black Forest Labs and Character.AI, plus Cursor, which competes with Claude Code [4].

Horowitz spoke inside a segment in which co-host David Sacks accused Anthropic of running a "scare campaign" to get the government to stifle competition, according to a caption index of the episode [3]. Horowitz did not name the firm that ran the test; he said only that it was "for the firm" [1].

What happened with the Claude Code leak?

Anthropic accidentally published internal Claude Code source code in an npm package on 31 March 2026 [5][6]. BleepingComputer reported a 60 MB source-map file from which about 1,900 files and 500,000 lines of code could be reconstructed [6].

Anthropic said then that it was a release packaging issue caused by human error, not a security breach, and that no customer data or credentials were exposed [5]. Posts on X quoting the full episode say Horowitz pointed to that leak as evidence; the clip does not include that part.

The AI Pin also covered another Horowitz remark from the same episode, that Amazon's compute is sold out beyond 2028 (theaipin.com/trends/horowitz-says-amazon-compute-is-sold-out-beyond-2028).

Our read

Opinion: An agent that finds and exploits a SQL injection on request is a real dual-use problem worth answering. But a one-sided anecdote from an investor backing Anthropic's rivals is not evidence of negligence, and Anthropic could settle it by saying what it told the firm.

What we could not verify

• The pet-shop test itself: no date, setup, prompt, data type or record of the call to Anthropic has been published.
• Horowitz did not name the firm; a16z is the likely but unconfirmed reading.
• The episode timestamp (about 72:37 to 73:45) comes from third-party caption indexes; YouTube blocked our direct check.
• Whether Horowitz cited the Claude Code leak in the full episode; we saw only posts on X quoting it.
• Whether a16z holds Anthropic shares through other vehicles; its published list does not show Anthropic, and one secondary source claims otherwise.