AI startup claims it hacked Anthropic in 12 hours, gains access to employee details and internal messages
A cybersecurity startup said it accessed some Anthropic employee information and internal messages through weaknesses in enterprise software.
TL;DR
- TechRadar reported that the startup Cathedral claimed it accessed employee details and internal messages after spending about 12 hours probing Anthropic’s systems.
- TechRadar, citing the Washington Post, said the startup reported that it did not access Anthropic source code or model weights.
- India Today separately reported Cathedral’s claim of access to employee information and internal messages.
TechRadar reported that the startup Cathedral claimed it accessed employee details and internal messages after spending about 12 hours probing Anthropic’s systems. The claim concerns access through ordinary enterprise software rather than a demonstrated compromise of Anthropic’s model weights. [1] [2]
TechRadar, citing the Washington Post, said the startup reported that it did not access Anthropic source code or model weights. The account describes the activity as a security test and says Anthropic detected some of it and took containment measures. [1] [2]
India Today separately reported Cathedral’s claim of access to employee information and internal messages. Both accounts attribute the central description to the startup; the available coverage does not independently establish the full scope of the access. [1] [2]
Anthropic’s system logs, the exact data viewed and whether any information was copied or misused remain unclear in the cited reporting. The company’s response and containment actions are reported, but the public accounts do not provide a complete incident record. [1] [2]
Why it matters
AI companies’ security depends on the enterprise systems around their models as well as the models themselves. The reporting is a company claim and does not establish source-code or model-weight theft.
Editor's note
The central access claim comes from Cathedral and remains incompletely corroborated; coverage says source code and model weights were not accessed.