Anthropic announced it identified 5,500 verified vulnerabilities between April and October 2026, while its partner Glasswing reported discovering over 129,000 vulnerabilities during the April to July period of the same year. The findings were shared in a Techmeme newsletter update on October 6, 2026, highlighting ongoing efforts to improve AI system security through collaborative vulnerability discovery. Both organizations emphasized the importance of third-party validation in strengthening model safety before public deployment.
The vulnerability discoveries stem from structured testing programs where external researchers and internal teams probe AI models for weaknesses that could lead to harmful outputs or unintended behaviors. Anthropic’s 5,500 figure represents issues confirmed as reproducible and impactful after initial screening, while Glasswing’s larger count reflects a broader initial sweep that included potential flags requiring further verification. The collaboration allows for cross-checking results, reducing false positives and increasing confidence in the severity of identified risks. This approach aims to catch flaws early in development cycles, particularly as models grow in size and complexity.
Anthropic uses a multi-stage process where potential issues are first flagged by automated tools or human testers, then subjected to rigorous validation to confirm they pose real risks under realistic conditions. Only after passing checks for reproducibility, severity, and exploitability are vulnerabilities counted in the final tally. Glasswing’s role involves independent testing using complementary methodologies, helping to uncover blind spots in internal assessments. Their April-July data included findings from red-team exercises focused on prompt injection, jailbreaking attempts, and bias amplification in conversational scenarios. Both firms stress that transparency about limitations and ongoing challenges is key to building trust in AI safety practices.
While high vulnerability counts might seem alarming, experts note that finding and fixing issues before release is a sign of maturing safety protocols rather than inherent instability. The volume of discoveries reflects increased scrutiny as AI systems are deployed in more sensitive domains like healthcare, finance, and education. Anthropic stated that each verified vulnerability leads to a patch or mitigation strategy, with critical flaws prioritized for immediate remediation. Glasswing added that their partnership model allows for faster iteration, as insights from one team’s testing inform the other’s focus areas. The ultimate goal is not zero vulnerabilities—which may be unattainable—but ensuring that known risks are managed effectively and do not reach end users.
What counts as a verified vulnerability in this context? A verified vulnerability is a flaw that has been confirmed to be reproducible, exploitable under defined conditions, and capable of causing harmful or unintended behavior in an AI model after rigorous testing.
Why does Glasswing’s number appear much higher than Anthropic’s? Glasswing’s count includes initial findings from broader screening efforts between April and July, many of which required further validation; Anthropic’s figure reflects only those issues that passed full verification checks by October.
How are these vulnerabilities addressed once discovered? Each verified vulnerability triggers a response that may involve retraining components, adjusting safety filters, or updating model architecture, with fixes deployed before public release or through patches for existing systems.