Sep 1, 2026 · 6 min
Claude session theft exposes AI access-control gaps
Claude sessions hijacked, AI jolts global finance, agents get too many keys
A hijacked browser session can bypass passwords and two-factor authentication, while broader AI risks extend from individual accounts to financial stability.
- 1Infostealer malware can let attackers hijack active Claude sessions without stealing passwords or bypassing two-factor authentication.
- 2Anthropic responded by signing out affected users, removing payment methods, and refunding unauthorized charges.
- 3The briefing connects AI security failures to wider concerns about financial stability and excessive agent permissions.
Don't miss
The clearest warning is that attackers can use a hijacked Claude browser session without needing the account password or two-factor authentication.
The brief
Sarah Lane opens with Anthropic’s warning that infostealer malware can hijack active Claude browser sessions, giving attackers access without passwords or two-factor authentication.
The practical fallout is immediate: attackers can consume paid Claude usage, while Anthropic signs out affected users, removes payment methods, and refunds unauthorized charges.
The briefing widens the lens to AI’s effect on financial stability, including concerns raised around the Bank of England, Andrew Bailey, and the G20.
A second security problem is excessive permission: AI agents often receive more keys and access than necessary, increasing the consequences when controls fail.
The episode’s central lesson is straightforward: AI adoption needs tighter access controls and incident response, not just confidence in the model itself.