Claude Sonnet 4 profile

    Claude Sonnet 4 — security & risk dashboard

    Known weaknesses, safety and security advisories, and documented real-world incidents for Claude Sonnet 4 — collected automatically and re-checked every few hours, so teams building on it can see what to design around.

    Last checked 25 Oct 2026

    Ask the Advisor about this model
    Known flaws

    5

    Security risks

    4

    Real-world incidents

    0

    Advisories logged

    0

    Security, safety & abuse risks

    • Anthropic has documented Claude Code misuse for data theft and extortion; this is application-level evidence, not confirmation of a particular Sonnet checkpoint.
    • Malicious actors can disguise harmful requests as legitimate security work to evade safeguards.
    • Broad filesystem, network and credential access can magnify the consequences of unsafe agent actions.
    • Untrusted documents, websites and tool outputs can carry prompt-injection instructions into agentic workflows.

    Known flaws & limitations

    • Published GPQA Diamond and MMMU results are below Claude Sonnet 3.7’s results, so benchmark gains are not uniform.
    • The published Terminal-Bench result shows substantial remaining limitations on complex terminal workflows.
    • Tool-assisted performance varies by domain; strong coding results do not establish reliable autonomous operation.
    • The 1M-token context option has eligibility requirements and premium pricing above 200K input tokens.
    • Generated code, factual claims and tool-derived conclusions still require independent validation.

    Security advisories & risk updates

    No risk advisories logged since this profile started being tracked. New ones appear here automatically.

    What this means if you are building on it

    • Unsupervised agents with unrestricted production, credential or financial access.
    • High-stakes decisions without qualified human review.
    • Workflows requiring consistently reliable autonomous terminal operation.
    • Applications that treat untrusted retrieved content as authoritative instructions.

    Everything here comes from published documentation, advisories and reporting, with links to the original sources. It is not a complete vulnerability database — always check the provider's own security notices before shipping.