TEL AVIV, Israel, Oct. 09, 2024 (GLOBE NEWSWIRE) — Pillar Security, a company specializing in GenAI security solutions, unveiled the industry’s inaugural "State of Attacks on GenAI" research report today. This report is a product of comprehensive analysis encompassing over 2,000 real-world AI applications. In stark contrast to previous conjecture and abstract risk assessments, this data-driven study relies on Pillar’s telemetry data extracted from live data interactions within operational AI applications spanning the last three months.
According to the report, several key revelations emerged:
-
High Success Rate of Data Theft: The research shows that 90% of successful attacks led to the compromise of sensitive data.
-
Alarming Bypass Rate: Approximately 20% of attempts involving jailbreak attacks managed to circumvent GenAI application safeguards.
-
Rapid Attack Execution: Adversaries typically took just 42 seconds on average to carry out an attack.
-
Minimal Interaction Required: Attackers needed only around five interactions with GenAI applications to execute successful attacks.
-
Widespread Vulnerabilities: The report underscores the pressing necessity for robust security measures as attacks exploited vulnerabilities at every stage of engagement with GenAI systems.
- Increase in Frequency and Complexity: The study highlights a noticeable rise in the frequency and complexity of prompt injection attacks. Attackers are employing more sophisticated tactics and persistently trying to evade safeguards as time progresses.
CEO and co-founder of Pillar Security, Dor Sarig, expressed, "The widespread integration of GenAI in businesses has opened new frontiers in cybersecurity. Our report surpasses theoretical concerns and, for the first time, sheds light on actual attacks transpiring in the wild. This offers organizations practical insights to strengthen their GenAI security framework."
The report uncovers numerous other noteworthy insights, including:
-
Top Jailbreak Techniques: These include directing AI systems to ignore initial programming instructions and encoding malicious prompts in Base64 to elude security filters.
-
Primary Attacker Motivations: These include theft of sensitive data, proprietary business information, personal identifiable information (PII), and sidestepping content filters to disseminate disinformation, hate speech, phishing messages, and malicious code.
-
Curated Attack List: A detailed examination of top attacks observed in real-world operational AI apps.
- Future Projections for 2025: Pillar foresees the transition from chatbots to copilots and self-governing agents, alongside the ubiquitous use of small, locally deployed AI models. While this new phase in AI adoption democratically expands access, it also broadens the attack surface, introducing additional security complexities for organizations.
Sarig further added, "As we advance towards AI agents capable of executing intricate tasks and decision-making, the security arena grows increasingly convoluted. Organizations must brace themselves for a surge in AI-targeted attacks by implementing specialized red-teaming exercises and adopting a ‘secure by design’ strategy in their GenAI development process."
The report emphasizes the inadequacy of traditional static security protocols in light of evolving AI threats. Jason Harrison, CRO of Pillar Security, stressed, "Static controls can no longer suffice in this dynamic AI-driven realm. Organizations need to invest in AI security solutions capable of predicting and responding to emerging threats in real-time while upholding their governance and cybersecurity policies."
For more information on AI Security, visit https://www.pillar.security/resources/buyer-guide. To schedule a demo, visit https://www.pillar.security/get-a-demo.
About Pillar Security:
Pillar Security is a leading provider of a cohesive platform to safeguard the entire AI lifecycle, from development to production to usage. The platform seamlessly integrates with existing controls and workflows, offering proprietary risk detection models, comprehensive visibility, adaptive runtime protection, robust governance features, and cutting-edge adversarial resistance. Pillar’s detection and evaluation engines continuously refine by training on extensive datasets of real-world AI app interactions, ensuring the highest accuracy and precision in identifying AI-related risks.