Top 5 AI Security Trends Discussed at the Confidential Computing Summit 2024
The 2-day conference featured discussions with the smartest minds in confidential computing and privacy-preserving (generative) AI.
2024 CoCO Summit
In June, I had the pleasure to both attend and speak at the 2024 CoCo Summit in San Franciso. My talk entitled, “Strategies for Effectively Deploying Trustworthy Generative AI Solutions”, was a bit generalized for this audience, as confidential computing solves only one aspect of LLM security, specifically on deployment.
Nonetheless, I received great feedback on my talk from the audience members, namely how impressed they were with the comprehensiveness of the Enkrypt AI platform. They appreciated its benchmarked and dynamic Red Teaming, Alignment, Guardrails, and continuous Monitoring capabilities. All of which can be done simultaneously in the platform.
We are proud of building a product that can (among other things):
- Detect both security risks (jailbreak, malware, leakage) and model risks (toxicity, bias and hallucinations), and
- Evaluate AI systems against operational and reputational risks throughout development and deployment.
The rest of the conference was filled with presentations from industry luminaries representing Microsoft, Nvidia, Google, and others.
Top 5 AI Security Trends
Here are the top trends I came away with after digesting the jam-packed content:
- Internal threat actors are increasing, so protecting LLM IP is becoming critical. And in some cases, of national security importance. Jason Clinton, CISO, at Anthropic essentially made this point in his presentation.
- The technology for confidential computing for Generative AI is not yet mature – confidential GPUs are a year away.
- Despite their infancy, use cases for confidential computing are starting to pick up steam. One example is to port the workloads (AI training, data processing) into confidential computing.
- Challenges abound at the CPU-GPU communication level when it comes to confidentiality.
- There is an obvious need to provide responsible and secure Generative AI. Threat actors know AI applications are currently an easy and profit-rich target to exploit.
We look forward to attending next year’s event, as interest will only grow in this industry.
Frequently Asked Questions
AI security protects generative AI systems from jailbreaks, data leakage, malware, and model risks like toxicity and hallucinations throughout development and deployment. Threat actors actively target AI applications as easy, profit-rich attack surfaces.
- Detects security risks: jailbreaks, malware, IP leakage across LLMs
- Identifies model risks: toxicity, bias, hallucinations in real-time
- Evaluates systems against operational and reputational risks continuously
Protecting LLM IP requires confidential computing, policy-based guardrails, and continuous monitoring to detect insider threats and unauthorized access. Internal threat actors are increasing, making IP protection critical and sometimes a national security concern.
- Deploy confidential computing to isolate AI training and data processing workloads
- Implement runtime guardrails that block unauthorized data exfiltration attempts
- Monitor access logs and model behavior for anomalies indicating insider compromise
AI security testing (red teaming) identifies vulnerabilities before deployment across 300+ risk categories, while continuous monitoring detects threats and drift in production. Enkrypt AI performs both simultaneously within a single platform.
- Red teaming: benchmarked, dynamic testing against jailbreaks and prompt injection
- Continuous monitoring: real-time detection of security and model risks post-deployment
- Alignment evaluation: ensures model behavior matches safety and compliance policies
Enkrypt AI is the only platform that combines benchmarked red teaming, alignment evaluation, policy-based guardrails, and continuous monitoring in a single system. Attendees at the 2024 Confidential Computing Summit praised its comprehensive approach to trustworthy AI deployment.
- Benchmarks and tests against 300+ red-teaming risk categories automatically
- Enforces security and compliance policies with real-time guardrails at runtime
- Reduces manual compliance effort by up to 90% across NIST, MITRE, and EU AI Act frameworks
Enkrypt AI's red teaming and monitoring capabilities address the deployment security gaps confidential computing alone cannot cover. Book a demo to see how it detects jailbreaks, leakage, and model risks across your LLM pipeline, or start a free trial today.
.avif)
.jpg)
.svg){kind=small}
