Anthropic Opens Data Leak Probe After Second Claude Outage in a Week

On the afternoon of June 5, Anthropic's Claude platform went offline for roughly three hours, the second significant service disruption in less than a week. When service resumed, a subset of users began posting on social media claiming they had received responses that appeared to belong to other people's sessions. Anthropic said Friday it is investigating those claims, though the company has not confirmed that any customer data was exposed.

The disruption began at 15:08 UTC when Anthropic's status page flagged elevated errors across several Claude models. Within minutes, problems had spread across claude.ai, the Claude API, Claude Code, and Claude Cowork. Recovery was staggered over the next three hours: Opus 4.6 returned to normal at 15:25 UTC, Sonnet 4.6 at 16:23 UTC, and Opus 4.8 at 16:59 UTC. Full service restoration was confirmed at 18:27 UTC. Anthropic's engineering team attributed the disruption to infrastructure issues rather than a security breach.

The Data Exposure Question

The more sensitive part of the story emerged afterward. A thread on X drew attention to posts from users who said Claude had returned coherent, on-topic responses that clearly originated from a different user's conversation. The posts were specific enough to prompt Cybernews to contact Anthropic directly for a response. Anthropic confirmed it is now investigating the claims.

Anthropic's public statement was careful and narrow. The company said it had not seen reports or evidence confirming a data leak but emphasized it takes any such claim seriously. Investigators are specifically examining whether Claude's inference infrastructure, under the strain of the recovery sequence, could have served cached or in-flight completions from one session to another. That is a known failure mode in distributed systems when request routing breaks down during degraded states.

"We haven't seen any other reports or evidence of customer data being leaked but take any reports of this very seriously and are investigating." Anthropic statement to Cybernews, June 6, 2026

A Pattern of Disruptions

The timing compounds the reputational exposure. The June 5 outage was not an isolated event. On June 2, the day after Anthropic filed its confidential IPO prospectus with the SEC, Claude went offline in a separate major disruption that affected services globally for several hours. That outage drew attention partly because of its timing: analysts noted the irony of a company pursuing a near-trillion-dollar public listing experiencing a high-profile service failure within 24 hours of announcing the filing. Then, on June 3, Claude Code services degraded for more than three hours overnight, hitting developers running long-horizon agent jobs. The June 5 outage was the third incident in four days.

This pattern has been building. Similar elevated-error incidents affecting Opus models have occurred several times over the past few months. Enterprises that have moved significant workloads onto Claude, including firms using it for agentic trading workflows and large-scale code generation, are beginning to factor reliability into their AI vendor assessments more explicitly than they did a year ago.

Enterprise Exposure and the Privacy Angle

If confirmed, a cross-session data leak would be qualitatively different from a plain service outage. Enterprises running Claude under compliance and data governance frameworks have contractual obligations around data handling that a pure availability failure does not trigger. A scenario where one customer's prompts are visible to another would implicate data protection laws in multiple jurisdictions, including the EU's GDPR, regardless of whether the exposed content was sensitive. Healthcare organizations, financial institutions, and legal firms using Claude would each face separate notification and investigation obligations under sector-specific regulations.

Anthropic has built a business case around enterprise trust. Its compliance integrations, security posture commitments, and participation in frameworks like the UK AI Safety Institute's testing program are all part of that pitch. The data leak investigation, however preliminary, lands at a sensitive moment. The company is in the process of converting its private market valuation into a public one, and enterprise customers will be watching how it handles this incident as a signal of institutional maturity.

Anthropic has not provided a timeline for completing its investigation or stated whether it plans to notify potentially affected customers before the findings are finalized. The company has not announced any changes to its incident response procedures as a result of this week's cluster of disruptions. For now, the clearest public commitment is the statement it gave Cybernews: the investigation is open.