Get the weekly research letter
Companies change their terms quietly. We read every version and catch what actually changed. One email a week on the changes that matter and what they mean. No account.
The document states that Opus 4.8 exhibits weaker prompt injection robustness than its predecessor Opus 4.7 in some agentic contexts, and that applied safeguards partially close this gap. Anthropic conducted a one-week live bug bounty competition to assess robustness against adaptive attackers across coding, computer use, and browser use surfaces.
This analysis describes what Anthropic's agreement states, permits, or reserves. It does not constitute a legal determination about enforceability. Regulatory applicability and practical outcomes may vary by jurisdiction, enforcement context, and individual circumstances. Read our methodology
This provision establishes that a production model deployed in agentic contexts has disclosed prompt injection vulnerabilities relative to its predecessor, with risk mitigation relying on applied safeguards rather than capability-level robustness improvements. Organizations deploying Opus 4.8 in autonomous or semi-autonomous agentic pipelines should assess this disclosure in the context of their specific deployment architectures.
Under these terms, users and developers deploying Opus 4.8 in agentic contexts, including Claude Code and computer use applications, operate under disclosed prompt injection risk that the document states is partially but not fully mitigated by applied safeguards. The bug bounty program provides a mechanism for security researchers to report discovered vulnerabilities.
Cross-platform context
See how other platforms handle Agentic Safety and Prompt Injection Vulnerability Disclosure and similar clauses.
Compare across platforms →Monitoring
Anthropic has changed this document before.
Receive same-day alerts, structured change summaries, and monitoring for up to 25 platforms.
"Although it shows improvements in some areas (such as refusing malicious requests), we found Opus 4.8 to be somewhat less robust than Opus 4.7 in several agentic contexts (such as vulnerability to prompt injection attacks). However, the application of our safeguards closes the gap between the models in practice. We report the results of our first one-week live bug bounty for prompt injection.Excerpt from Anthropic's Claude Opus 4.8 System Card
(1) REGULATORY LANDSCAPE: Prompt injection vulnerabilities in agentic AI systems are relevant to emerging regulatory guidance on AI system security, including the EU AI Act's security and robustness requirements for high-risk AI systems, and NIST AI Risk Management Framework recommendations. The FTC's unfair or deceptive practices authority may also engage disclosures about known security vulnerabilities in deployed AI systems. (2) GOVERNANCE EXPOSURE: High. The document discloses that a production model in agentic deployment contexts has lower prompt injection robustness than its predecessor, with safeguards rather than model-level improvements as the primary mitigation. Organizations deploying Opus 4.8 in agentic workflows with access to sensitive data, financial systems, or operational infrastructure face heightened exposure to prompt injection-based misuse. (3) JURISDICTION FLAGS: EU/EEA deployments under the AI Act's security requirements for high-risk AI systems, and US deployments in regulated sectors including finance and healthcare, create heightened exposure for agentic applications subject to prompt injection risk. The one-week bug bounty duration may be assessed by external reviewers as a limited timeframe for comprehensive vulnerability discovery. (4) CONTRACT AND VENDOR IMPLICATIONS: B2B customers deploying Opus 4.8 in agentic pipelines should assess whether their service agreements with Anthropic address liability allocation for losses arising from prompt injection attacks, and whether the disclosed vulnerability affects their own customer-facing security representations. Vendors building on Claude Code or computer use APIs should review their own terms of service in light of this disclosure. (5) COMPLIANCE CONSIDERATIONS: Compliance teams should assess whether the agentic safety disclosure requires updating their internal AI risk registers, incident response plans, or vendor security assessments. Organizations in regulated sectors should evaluate whether prompt injection vulnerability in deployed agentic systems requires notification to regulators or customers under applicable security incident disclosure frameworks.
This provision establishes that a production model deployed in agentic contexts has disclosed prompt injection vulnerabilities relative to its predecessor, with risk mitigation relying on applied safeguards rather than capability-level robustness improvements. Organizations deploying Opus 4.8 in autonomous or semi-autonomous agentic pipelines should assess this disclosure in the context of their specific deployment architectures.
Under these terms, users and developers deploying Opus 4.8 in agentic contexts, including Claude Code and computer use applications, operate under disclosed prompt injection risk that the document states is partially but not fully mitigated by applied safeguards. The bug bounty program provides a mechanism for security researchers to report discovered vulnerabilities.
No. ConductAtlas is an independent monitoring service. We are not affiliated with, endorsed by, or sponsored by Anthropic.