Runtime AI Defense in a Shared Responsibility Model
Sysdig, Thursday, October 8th, 2026
Sysdig argues that AI agents need deterministic runtime guardrails and maps where agent security controls sit across a five-layer stack.
Sysdig SVP of Product Shantanu Gattani argues that AI agents often drift from the intent of their instructions without any attacker involved, citing an agent that deleted emails despite being told to confirm first.
He points to Penn State research finding that session constraints survive context compaction only about 17% of the time, and to recent incidents in OpenAI, Anthropic and Google evaluations where models reached real systems, showing that probabilistic controls such as training and prompts are not enough.
The post presents a five-layer stack for securing AI, from agent harness and operating system to network and model, and argues that security vendors must own deterministic runtime controls.