Anthropic Auto-Logs-Out Claude Users After Credential Theft

Automatic sign‑out follows credential theft

Anthropic began terminating active Claude sessions when it detected that an infostealer had harvested login cookies from users’ machines.¹²³⁴⁵ The move came after security researchers reported that the malware captured active sessions and could be used to hijack accounts.^12673485 Anthropic responded by adding a server‑side check that forces a logout for any session matching known compromised tokens.¹²⁷³⁴⁵ The rollout was announced on Engadget¹, which cited the malware’s ability to pull active Claude login sessions from PCs.¹ Anthropic did not disclose the exact detection method, but the automatic sign‑out is now part of the service’s default behavior.¹⁷³⁵

Claude’s broader security posture and the Mythos debate

Anthropic’s response to the credential theft sits alongside a longer‑running discussion about the security implications of its most advanced models. The company recently previewed Claude Mythos, a model built under Project Glasswing with strong coding and vulnerability‑analysis capabilities. Anthropic limited public access after internal tests showed the model could identify and exploit software weaknesses at a level that raised concern. Financial Times and Reuters both reported that regulators and banks in Asia are reassessing AI‑driven cyber risk in light of Mythos. The reports note that the model compresses the time between discovering a vulnerability and weaponizing it. Anthropic’s own statements acknowledge the trade‑off: powerful assistance for developers versus the potential for malicious automation.

Batch processing discounts and operational trade‑offs

In a separate development, Anthropic released a batch‑processing feature for Claude Code that offers a 50 % discount on non‑urgent jobs. The batch API accepts prompts that can wait roughly an hour and runs them on Claude Opus at half price. Users can route jobs through Anthropic’s API directly or via Google Cloud’s Vertex AI, which uses the same discount structure. The batch workflow requires an API key stored in a protected file and a manifest that lists every change before execution. The tool does not run background daemons; it polls the API only when a job’s status is queried. This design lowers resource overhead but adds a manual step for developers who need to monitor long‑running tasks.

Emerging tooling to police AI‑generated code

Security‑focused developers are deploying third‑party tools to mitigate the risk of malicious code generated by Claude or ChatGPT. A free Chrome extension called Secure AI Generated Code scans every code snippet displayed on claude.ai and chatgpt.com, flagging known security patterns in real time. The extension integrates directly with the web pages, offering instant feedback without requiring a separate IDE. Backplanes, another startup, offers a reporting layer called Spotlight that aggregates Claude Code and Codex sessions across an organization. Spotlight captures session telemetry, redacts credentials, and surfaces risk findings to security, engineering, and finance teams. The service aims to make invisible AI activity visible, allowing enterprises to enforce policy on credential exposure and external‑domain access.

What to watch

Anthropic’s automatic sign‑out is a reactive measure; the underlying credential‑theft vectors remain active. Watch for any follow‑up announcements from Anthropic about proactive token‑rotation or multi‑factor enforcement for Claude accounts. Track regulatory responses to Claude Mythos as governments evaluate whether advanced AI models need licensing or usage caps. Monitor adoption rates of the batch API discount, especially among teams that shift non‑critical workloads to lower‑cost processing. Finally, keep an eye on enterprise‑level tooling like Secure AI Generated Code and Backplanes, which could become de‑facto standards for AI‑driven development security.