# Anthropic AI development and security

> Developing situation record from CLSTR: https://clstr.news/situations/anthropic-ai-development-and-security
> Updated: 2026-10-26T11:18:58.000Z. Sources: 21. Developments: 2.

Anthropic is conducting research into training AI agents to perform complex negotiations, such as those involving contracts and salaries, by mimicking professional legal standards. The company is reportedly consulting with law firms to analyze how attorneys advocate for client interests and seek equitable outcomes. As autonomous agents are expected to handle increasing commercial and personal disputes, Anthropic is studying interaction norms between models, specifically determining when an agent should request human intervention and how to establish appropriate levels of skepticism to prevent dysfunctional behavior.

Experimental data from these studies indicates a strong correlation between an AI model’s capabilities and its negotiation success. In tests involving the exchange of goods via AI proxies, higher-performing models consistently secured better results for their users, exhibiting social dynamics similar to human negotiators.

In a separate development regarding the company’s security, the cybersecurity startup Cathedral used AI agents during an authorized testing program to access Anthropic’s internal systems in under 12 hours. The researchers exploited vulnerabilities in third-party administrative and sales software to access data including employee compensation, meeting transcripts, and Slack discussions regarding advanced AI models. 

Anthropic stated that while some sales, human resources, and IT systems were accessed, core AI infrastructure, including Claude customer conversations, model weights, and source code, remained secure. The company disabled the affected software until vendors provided fixes and reported the incident to federal authorities. While the breach raised national security concerns among government officials regarding the potential for state-sponsored actors to replicate such automated attacks, Anthropic characterized the implications as overstated. Cathedral was founded by former staffers of the Department of Government Efficiency (DOGE), and its co-founder previously served as the Pentagon’s chief data officer.

## Claims

- AI agents from Cathedral accessed Anthropic’s internal systems in less than 12 hours. (reported by 12 sources)
- The breach exploited vulnerabilities in third-party administrative and sales software. (reported by 11 sources)
- Researchers accessed employee salary details, job offers, and meeting transcripts. (reported by 11 sources)
- Anthropic confirmed that customer conversations and model weights were not accessed. (reported by 10 sources)
- The testing was conducted as part of an authorized bug bounty or security research program. (reported by 10 sources)
- Cathedral was founded by former staffers of the Department of Government Efficiency (DOGE). (reported by 6 sources)
- Anthropic reported the incident to federal authorities. (reported by 6 sources)
- Anthropic disabled the affected software until vendors confirmed fixes. (reported by 5 sources)

## Timeline

### 2026-10-26: Anthropic systems breached by AI agents in under 12 hours

AI agents from startup Cathedral breached Anthropic’s internal business systems in under 12 hours via third-party software vulnerabilities, accessing employee data but leaving core AI models untouched.

18 sources. https://clstr.news/cluster/anthropic-internal-systems-accessed-by-ai-powered-security-testing

### 2026-10-26: Anthropic trains AI agents to negotiate like lawyers

Anthropic is developing AI agents capable of negotiating contracts and disputes by integrating professional legal standards and studying human-like negotiation dynamics.

3 sources. https://clstr.news/cluster/anthropic-trains-ai-agents-to-negotiate-like-lawyers

---
Cite as: Anthropic AI development and security. CLSTR, https://clstr.news/situations/anthropic-ai-development-and-security