All Radar picks

X OPPORTUNITY RADAR / DISPATCH

The Frontier Lab Had an Office-Software Side Door

Researchers reached limited Anthropic sales, HR and IT systems through third-party software in under 12 hours. The model vault held; the enterprise stack did not.

Information Overload
2 min readSources
Editorial illustration of a glowing frontier-AI vault with a flimsy filing-cabinet side door marked only by office folders, while a red security probe enters through the business-software corridor.
AI-generated editorial illustration

Cybersecurity researchers working through Anthropic’s testing program reached parts of the company’s sales, human-resources and IT systems in under 12 hours by exploiting weaknesses in ordinary third-party business software. The Washington Post reports that the exposed material included meeting transcripts, salary and job-offer details, Slack discussions, business contacts and some basic user-account records.[1]

The scope matters. The researchers did not reach customer Claude prompts or responses, model weights or Anthropic source code. This was not a break into the model itself. It was a demonstration that an AI lab can spend lavishly on frontier-model security while still inheriting the much duller attack surface of the software used by every large office.[1]

Anthropic says its monitoring detected some of the activity before the report, that it cut access the same day, disabled affected software until vendors fixed the paths, and found no evidence of outside exploitation. Anthropic also says Cathedral overstated the national-security implications and did not fully follow the bug-bounty rules; Cathedral says it limited its access.[1]

That dispute should not hide the practical lesson. The most advanced system in the building can still sit behind an ecosystem of sales tools, meeting software and vendor integrations that were not built for the stakes now attached to it. Which connection gets audited first: the glamorous model endpoint or the boring app with an old token and a generous permission set?

FOLLOW THE SOURCES.

Checked Oct 26, 2026. Numbered references point to the evidence behind factual passages. Commentary is our own.

  1. Researchers penetrated Anthropic by exploiting its business software (opens in a new tab)The Washington PostBack to cited passage
AIcybersecurityenterprise softwaresupply-chain riskAnthropicCathedral
Back to the latest picks