Using eBPF Runtime Security to Protect Financial Systems and AI Agents
· 54:56 · Watch on YouTube ↗
Summary
Rashid Feroze discusses using eBPF-based runtime monitoring to detect suspicious behavior in financial services, its limitations against prompt injection, and practical steps for deployment and vulnerability response.
Key points
- eBPF-based sensors can monitor process execution, network connections and file activity, helping identify suspicious behavior such as a payment API launching a shell.
- Feroze described a compromised npm package that collected environment variables and secrets, archived them and attempted to upload them to a C2 server.
- Runtime controls may detect or block some consequences of prompt injection, but they do not reliably identify prompt injection or data exfiltration through authorized channels.
- Feroze recommends starting in detection mode, assessing false positives and selectively enabling blocking after observation; his team's deployment took about 90 days, not including a full transition to blocking.
- Managed cloud services may prevent customers from installing host agents, limiting direct visibility and control.
- Feroze recommends maintaining a searchable software bill of materials to identify affected services and using runtime security as a supplementary control while patching.
Article Details
- Topics
- [03:08] eBPF-based runtime monitoring and system-call visibility
- [07:27] AI agent security and prompt-injection limitations
- [11:15] Runtime detection of supply-chain attacks and exploitation
- [19:45] Detection rules, behavioral baselines and Falco
- [21:30] Deployment limitations in managed cloud services
- [33:31] Deployment planning, detection mode and selective blocking
- [39:19] Performance benchmarking, API inventory and egress visibility
- [50:50] Software bills of materials and vulnerability response