Skip to main content
Mintive is the new name for MintMCP.Read the announcement
Jiquan Ngiam

24 Prompt Injection Statistics (2027): Attack Rates, Success Rates and Defenses

Enterprise data reveals the scale of prompt injection threats, documented success rates, and the security controls that separate protected organizations from vulnerable ones

Prompt injection holds the #1 position in the OWASP Top 10 for LLM Applications, while Agent Goal Hijack (ASI01), which includes prompt injection techniques, ranks first in the OWASP Top 10 for Agentic Applications. Both frameworks identify instruction manipulation as a major security risk for enterprise AI systems. Valid prompt injection reports on bug bounty platforms increased 540% year-over-year, while 82% of organizations report using AI agents and only 44% have security policies in place. Mintive (formerly MintMCP) provides Mint Guard, which screens MCP gateway tool calls for prompt injection and sensitive data. In Enforcing mode, it blocks high-confidence prompt injection detections and records findings for security review. The statistics below document the attack landscape, success rates under controlled testing, and the governance gap that leaves enterprises exposed.

Key Takeaways​

  • Attack reports are exploding - Valid prompt injection reports increased 540% year-over-year on HackerOne, with system prompt leakage up 557%
  • Indirect attacks dominate - Indirect prompt injection achieved a 27.1% session-level success rate vs 5.7% for direct attacks (4.75x higher), requiring protection at the data layer
  • Adaptive attacks overcome defenses - Adaptive attacks succeeded above 90% against most of 12 recent prompt injection defenses tested, highlighting the need for defense-in-depth
  • Enterprises lack defenses - 82% use AI agents but only 44% have security policies; 97% of organizations experiencing AI-related security incidents lacked proper AI access controls
  • Data leakage is the primary outcome - 90% of successful attacks result in sensitive data exfiltration
  • Agent market approaches $236B by 2034 - The AI agents market is projected to reach $236.03 billion by 2034 at 45.82% CAGR, with one in three enterprises operating 500+ AI agents by 2028

Attack Volume and Industry Rankings​

1. Prompt Injection Ranks #1 in the OWASP Top 10 for LLM Applications​

The OWASP Gen AI Project identifies prompt injection as LLM01 in its 2025 and 2026 LLM application risk rankings, reflecting its importance as a security risk rather than a measured rate of exploitation in production deployments. This ranking reflects both the frequency of attacks and the severity of potential outcomes, including data exfiltration, unauthorized actions, and complete system compromise.

2. Agent Goal Hijack (ASI01) Ranks #1 in the OWASP Top 10 for Agentic Applications 2026​

The December 2025 OWASP release specifically addresses autonomous AI agents, with agent goal hijack through indirect prompt injection taking the top position. This validates that agentic systems face elevated risk compared to chat-based LLM deployments due to their ability to execute real actions.

3. Valid Prompt Injection Reports Increased 540% Year-Over-Year on HackerOne​

Bug bounty data from HackerOne's 2025 report shows an unprecedented surge in security researcher attention to prompt injection vulnerabilities. This growth rate far exceeds other vulnerability categories, reflecting both the proliferation of AI systems and the relative ease of discovering injection flaws.

4. System Prompt Leakage Reports Increased 557% Year-Over-Year on HackerOne​

The 2026 Security Research Report documents even faster growth in system prompt leakage findings from July 2025 to June 2026. Leaked system prompts expose internal instructions that attackers use to craft more effective injection attempts, turning a disclosure into a roadmap for exploitation.

5. 1,121 Customer Programs Included AI in Their Bug Bounty Scope in 2025, up 270% Year-Over-Year​

The expansion of AI-focused bug bounty programs indicates enterprises recognize AI security as a material risk requiring external validation. This growth creates a feedback loop where more programs generate more research, exposing more vulnerabilities.

Attack Success Rates Under Controlled Testing​

6. Indirect Prompt Injection Achieved a 27.1% Session-Level Success Rate vs 5.7% for Direct Attacks (4.75x Higher)​

The Red Teaming Challenge analyzed 1.8 million attacks and found indirect attacks through data sources vastly outperform direct chat-based attempts. This validates that organizations must secure their MCP connectors and data pipelines, not just their chat interfaces. Agent Monitor provides visibility into gateway activity and supported local agent activity, helping teams investigate where indirect attacks could enter across monitored environments.

7. Over 60,000 Prompt Injection Attacks Produced Policy Violations in a 1.8 Million Attack Study​

The same research documents the absolute volume of successful attacks against models without additional safeguards. This demonstrates that prompt injection is not a theoretical risk but a reliably exploitable vulnerability at scale.

8. Claude 3.5 Sonnet Agent Hijacking Success Increased From 11% to 81% With Adaptive Red-Team Attacks​

NIST CAISI testing revealed that standard evaluation methods underestimate real attacker capability. When evaluators used techniques matching actual adversary behavior, success rates jumped dramatically, indicating that static benchmarks provide false confidence.

9. Sophisticated Attackers Bypass Safeguards Approximately 50% of the Time With 10 Attempts on Best-Defended Models​

The International AI Safety Report 2026 establishes that even leading models with built-in defenses remain vulnerable to determined attackers with minimal persistence. No model-level defense alone provides sufficient protection.

10. Attack Success Rates Reached 50% to 84% Depending on Technique in a Dataset With Over 461,640 Submissions​

A Proofpoint-cited study shows wide variance in success rates based on attack technique, confirming that defenses must address multiple attack vectors simultaneously rather than optimizing for a single method.

11. Adaptive Attacks Succeeded Above 90% for Most of 12 Recent Prompt Injection Defenses Tested​

Research from Nasr et al. demonstrates that purpose-built attacks defeat most published defense mechanisms when evaluated properly. This underscores the need for defense-in-depth rather than reliance on any single mitigation technique.

12. Injections Optimized Against Gemini 2.0 Succeeded 92% of the Time on That Model​

Google DeepMind research shows near-total success when attackers optimize for a specific target. On the same evaluation set, adversarially trained Gemini 2.5 reduced attack success to 18%. The researchers cautioned that further adaptive testing revealed additional vulnerabilities, reinforcing the need for continuous evaluation alongside training-time and runtime defenses.

CVE Data and Real-World Exploitation​

13. 141 CVE Records in the NIST National Vulnerability Database Name Prompt Injection​

Analysis of the NIST NVD API as of October 2026 found prompt injection explicitly named in 141 vulnerability records, with 87 published in 2026 alone. This formal CVE recognition brings prompt injection into the same compliance and patch-management frameworks as traditional software vulnerabilities.

14. 105 of 141 Prompt Injection CVEs Were Rated High or Critical CVSS Base Severity​

The same analysis found 74% of prompt injection CVEs received High or Critical severity ratings, indicating these vulnerabilities routinely meet thresholds for mandatory remediation under most enterprise security policies.

15. Microsoft 365 Copilot EchoLeak Vulnerability (CVE-2025-32711) Received a CVSS Score of 9.3​

The Vectra AI assessment highlights this critical vulnerability in a widely deployed enterprise AI tool. High-severity CVEs in mainstream productivity tools demonstrate that prompt injection risk extends to organizations through their standard software stack.

16. GitHub Copilot Remote Code Execution Vulnerability (CVE-2025-53773) Received a CVSS Score of 9.6​

The same source documents a near-maximum severity rating for a coding assistant vulnerability. Organizations using AI coding tools face exposure through tool repositories and code context that attackers can manipulate.

17. Cursor IDE Vulnerability (CVE-2025-54135) Received a CVSS Score of 9.8​

The NIST NVD assigns a 9.8 CVSS v3.1 score to a Cursor vulnerability affecting versions before 1.3.9 that could enable remote code execution through prompt injection. Teams using multiple coding assistants need centralized governance rather than per-tool security approaches. Mintive's Agent Monitor setup for Cursor supports visibility into Cursor activity.

18. Adversaries Abused Legitimate GenAI Tools at More Than 90 Organizations in 2025​

The CrowdStrike 2026 Report documents real-world exploitation of enterprise AI deployments, moving prompt injection from research concern to operational threat. These attacks used injected prompts against tools already deployed and trusted within target organizations.

19. 90% of Successful Prompt Injection Attacks Resulted in Sensitive Data Leakage​

Pillar Security research establishes data exfiltration as the dominant outcome of successful attacks. This aligns prompt injection risk with data protection requirements including GDPR, HIPAA, and SOC 2 controls.

Enterprise Governance Gaps​

20. 82% of Organizations Use AI Agents, but Only 44% Have Security Policies in Place​

The SailPoint AI Agent Adoption Report documents the fundamental governance gap in enterprise AI deployment. Adoption far outpaces security infrastructure, creating exposure that grows with each new agent deployment. Mintive's MCP Gateway addresses this gap by enforcing authentication, authorization, and audit on every AI tool connection.

21. 73% of CISOs Are Very or Critically Concerned About AI Agent Risks, but Only 30% Have Mature Safeguards​

The State of AI Agent Security Report shows leadership awareness exceeds organizational readiness by a factor of 2.4x. This gap represents both a risk and an opportunity for security teams to drive governance investments.

IBM Data Breach Report data connects security incidents directly to control failures. Organizations without access governance face near-certain incident exposure as their AI usage scales.

23. 80% of Companies Say Their AI Agents Have Taken Unintended Actions​

The SailPoint report documents that autonomous behavior already produces unwanted outcomes in most organizations. Prompt injection provides attackers a reliable method to transform unintended actions into malicious ones.

24. 23% Reported Their AI Agents Have Been Tricked Into Revealing Access Credentials​

The same research shows credential exposure through social engineering of AI systems is not hypothetical. Nearly one in four organizations has experienced this specific attack outcome.

Market and Financial Impact​

The AI prompt security market grew from $1.51 billion in 2024 to $1.98 billion in 2025, representing 31.5% CAGR. Projections reach $5.87 billion by 2029 as enterprises recognize security investment as essential to AI deployment.

Organizations extensively using AI and automation in security operations reported $1.9 million lower average breach costs compared with organizations not using these technologies. With 40% of organizations experiencing AI agent breaches estimating losses between $1 million and $10 million, governance and security investment remain important considerations for enterprise AI deployments.

The AI agents market is projected to reach $236.03 billion by 2034 at 45.82% CAGR. By 2028, one in three enterprises will operate 500+ AI agents, making governance at scale a critical capability.

Building Effective Defenses​

Prompt injection defense requires multiple layers working together. No single control provides complete protection. Effective enterprise programs combine:

  • Runtime screening through tools like Mint Guard that inspect every tool call for injection patterns before execution
  • Per-agent identity via Agent Bundles that scope each agent's access to only required tools with independent credential rotation
  • Centralized audit through SIEM export that feeds AI activity into existing security operations
  • Custom policy enforcement via gateway middleware that runs customer-authored logic on every call

How Mintive Addresses Prompt Injection at Scale​

Mintive provides enterprise organizations with comprehensive protection against prompt injection through a defense-in-depth architecture that combines multiple security controls. The platform's design recognizes that no single layer fully prevents prompt injection, requiring runtime detection, identity-based access controls, centralized governance, and continuous audit.

Runtime Detection and Prevention​

Mint Guard screens MCP gateway tool calls for prompt injection patterns, credentials, PII, and harmful content before execution. In Enforcing mode, Mint Guard blocks high-confidence prompt injection detections and records each finding for security review. The service runs as managed detection policies that inspect both tool arguments and results, providing protection at the data layer where indirect attacks prove 4.75x more effective than direct chat-based attempts.

Identity-Based Access Control​

Agent Bundles provide each AI agent with its own identity and credential isolation. Each agent receives scoped access to only the tools required for its specific function, with independent credential rotation and audit trails. This architecture limits blast radius when an agent is compromised, preventing lateral movement across the organization's AI infrastructure. Per-agent credentials replace shared service accounts that create credential sprawl and attribution gaps.

Centralized Governance and Visibility​

Mintive's MCP Gateway governs data and tool connections through Virtual MCPs, while its Agent Gateway extends that foundation to first-class agent identities, scoped permissions, and auditable credentials. Agent Monitor adds visibility into supported agent activity including prompts, commands, and file access across Claude, Cursor, ChatGPT, Gemini, and Copilot. Organizations gain unified observability across heterogeneous AI deployments rather than managing security on a per-tool basis.

Gateway middleware enables customer-authored policies and DLP integrations that run on every call. This extensibility allows organizations to enforce custom business logic, integrate with existing security infrastructure, and adapt controls as new threats emerge. SIEM export feeds AI activity into existing security operations, enabling correlation with other enterprise telemetry and incident response workflows.

Enterprise Compliance and Attestation​

The platform is SOC 2 Type II audited and compliant with HIPAA standards, and Mintive signs BAAs for customers handling protected health information. All gateway traffic receives authentication, authorization, and audit enforcement. This compliance foundation addresses the gap documented in enterprise adoption statistics, where 82% of organizations use AI agents but only 44% have security policies in place. Mintive provides the control layer enterprises need to deploy AI at scale while meeting regulatory requirements for data protection, access controls, and audit trails.

Frequently Asked Questions​

What is prompt injection and why is it the top AI security threat?​

Prompt injection occurs when attackers manipulate AI system inputs to override intended behavior, execute unauthorized actions, or extract sensitive data. It ranks #1 as LLM01 in OWASP's Top 10 for LLM Applications, while Agent Goal Hijack (ASI01), which includes indirect prompt injection techniques, ranks first in the OWASP Top 10 for Agentic Applications. Both rankings highlight the severity of instruction manipulation risks, including data exfiltration, unauthorized actions, and system compromise.

How do indirect and direct prompt injection attacks differ in success rates?​

In the AI Agent Red Teaming Challenge, direct attacks through chat interfaces achieved a 5.7% session-level success rate, compared with 27.1% for indirect attacks through external data sources. Indirect attacks prove 4.75x more effective because they inject malicious content through trusted data channels the AI processes automatically, bypassing user-facing input validation.

What compliance requirements address prompt injection in regulated industries?​

Prompt injection creates exposure under multiple security and compliance frameworks. HIPAA safeguards apply when covered entities and business associates handle protected health information, while SOC 2 assessments evaluate relevant security controls within the organization's defined system scope. Prompt injection CVEs can inform vulnerability assessment and remediation priorities, but CVE assignment alone does not impose universal patching requirements. The article identifies 141 prompt injection CVE records in the NIST National Vulnerability Database as of October 2026, with 105 rated High or Critical severity, underscoring the importance of vulnerability management and access controls.

How does Mintive's platform specifically address prompt injection risks?​

Mint Guard provides managed detection policies that screen every MCP tool call for prompt injection, credentials, PII, and harmful content. Agent Monitor captures supported agent activity including prompts, commands, and file access. Gateway middleware enables customer-authored policies and DLP integrations. The platform is SOC 2 Type II audited and compliant with HIPAA standards, and Mintive signs BAAs for customers handling protected health information.

Can prompt injection be fully prevented or is mitigation the primary strategy?​

Current evidence shows no model-level defense fully prevents prompt injection. Adaptive attacks succeed above 90% against most published defenses. Effective protection combines training-time hardening, runtime detection, access controls, and audit. The goal is reducing successful attacks while maintaining visibility and incident response capability for attempts that reach production systems.