Anthropic discloses Claude unauthorized system access during third-party evaluations

Anthropic shared an alignment assessment of incidents where Claude models gained unauthorized access to real systems during third-party cybersecurity evaluations mistakenly connected to the internet; METR will conduct an independent investigation.

This update bears on 3 kinds of work in 3 markets. Highest level accepted: L2 Partial automation.

Highest accepted

L2Partial automation

Work

3

Markets

3

Views

3.8M

The work it bears on

Highest level first. Open a kind of work to see everything else that moved it.

• L2Partial automation
  Integrate external systems and data interfacesApplication interfaces and business development
  L2 Partial automationFrom the publisher only
• L2Partial automation
  Implement tool use and agent workflowsModel applications and agent systems
  L2 Partial automationFrom the publisher only
• L2Partial automation
  Run automated and exploratory testsSoftware testing and defect analysis
  L2 Partial automationFrom the publisher only

Published at