Anthropic allegedly lowered AI safeguards for big-spend contracts, former employee says

Why it matters

If true, the trade‑off between safety and revenue could expose enterprise customers to higher risk of AI misuse and trigger regulatory scrutiny.

Claim from a former Anthropic employee

Former Anthropic engineer Adi Baradwaj posted a three‑tweet thread on X on July 26, 2026, alleging that the AI startup "lowered safeguards in exchange for large committed‑spend contracts." Baradwaj wrote that he observed multiple instances where Anthropic reduced safety restrictions on its Claude models for customers who signed sizable usage agreements. The thread can be read in full here.

Baradwaj also asserted that "most black‑hat hackers use standard Claude Code/Codex subscriptions for hacking, while white‑hat defenders ..." and quoted a second participant that "safeguards aren't sufficient to stop a moderately dedicated actor." No other Anthropic employee or spokesperson has corroborated these statements, and the company has not issued a public response.

Anthropic's publicly stated safety posture

Anthropic, founded in 2020 by former OpenAI researchers, has built its reputation on a "Constitutional AI" approach that embeds ethical guidelines directly into the model's decision‑making process. The company markets its Claude series as a safer alternative to competing large language models, emphasizing "robust guardrails" and an "iterative safety‑first development cycle" on its official site.

In 2023, Anthropic released an AI safety & policy white paper that described a layered system of content filters, refusal mechanisms, and a continuously updated safety constitution. The paper acknowledged the tension between model capability and safety, noting that "higher‑risk deployments may require custom safety configurations negotiated with the customer."

Baradwaj's allegation suggests that, in practice, that negotiation may have crossed a line from customization to outright reduction of safeguards, presumably to satisfy commercial pressure from large enterprise contracts.

Enterprise contracts and the economics of safety

Anthropic's revenue model relies heavily on usage‑based pricing for its API, with discount tiers for enterprises that commit to multi‑year, high‑volume spend. Industry analysts estimate that a single enterprise contract can reach tens of millions of dollars annually, especially when the client integrates Claude into customer‑facing applications, internal knowledge bases, or code‑generation pipelines.

If the claim is accurate, the alleged trade‑off would mean that Anthropic has prioritized revenue from these contracts over the uniform safety standards it promotes to the broader market. Such a move could create a two‑tier safety regime: a stricter baseline for low‑spend users and a relaxed set of guardrails for high‑value customers.

The broader AI‑security landscape

Baradwaj's observation that "most black‑hat hackers use standard Claude Code/Codex subscriptions" aligns with recent security research showing that publicly available API keys are frequently repurposed for illicit activities, from automated phishing to code injection. A 2025 report by the Center for AI Safety documented dozens of jailbreak attempts targeting Claude, often leveraging the model's code‑generation endpoint to produce malicious scripts.

Conversely, the same report noted that ethical security researchers and red‑teamers frequently rely on the same subscription tiers to test defenses, highlighting a blurred line between legitimate security work and abuse.

The claim that "safeguards aren't sufficient to stop a moderately dedicated actor" underscores a longstanding tension in AI deployment: no set of filters can fully prevent a determined user from coaxing a model into disallowed behavior. However, the core of the allegation is that Anthropic may have intentionally weakened those filters for a subset of customers.

Potential implications for customers and the market

If Anthropic indeed offers reduced‑safeguard contracts, enterprise buyers could face higher exposure to model misuse, regulatory scrutiny, and reputational risk. Organizations that rely on Claude for customer‑service chatbots, code assistance, or content generation might need to implement additional downstream monitoring, increasing operational overhead.

Regulators in the EU and US have begun drafting AI‑risk frameworks that could treat safety‑guard reductions as a material change in risk profile, potentially triggering reporting obligations under upcoming AI Act provisions. Investors may also reassess the valuation premium Anthropic commands as a safety‑focused brand.

Verification challenges and next steps

Baradwaj's thread provides no direct evidence—no contract excerpts, internal memos, or screenshots of altered model configurations. The lack of corroborating testimony from other Anthropic staff, customers, or third‑party audits means the claim remains unverified.

Industry observers will likely seek independent verification through code‑level analysis of Claude's API responses under different contractual conditions, or via whistleblower disclosures. Until such evidence emerges, the story stands as a serious allegation that adds to ongoing debates about the commercial pressures shaping AI safety.

Context for readers

Anthropic’s approach to safety has been a differentiator in a crowded generative‑AI market dominated by OpenAI, Google, and Microsoft. The company’s recent $4 billion Series C round in 2024, led by Andreessen Horowitz and Sequoia, emphasized its "long‑term safety focus" as a core value proposition. Baradwaj’s claim, if true, would represent a stark contrast to that narrative and could influence how investors, enterprise buyers, and regulators evaluate the trade‑offs between performance, cost, and safety in AI services.