Tamnoon's four-agent system on Bedrock cut critical cloud exposures 90% in 90 days
AWS ML Blog · rss · 2026-10-14
An AWS blog details how security firm Tamnoon built Tami, an agentic cloud remediation orchestrator on Amazon Bedrock, to autonomously process the flood of findings from CNAPPs and cloud-native tools. Results: over 90% reduction in critical cloud exposure in 90 days with zero production issues.
Four specialized agents form the core:
- Investigation agent: reasons over context trained on millions of real customer environments, filtered by an LLM-as-a-judge validation layer, producing read-only investigation plans
- Safety-scoring agent: builds an evidence chain per asset and assigns a conservative SAFE / RISKY / AWAITING DATA Remediation Confidence Indicator (RCI), downgrading when in doubt
- Remediation agent: activates only on SAFE findings; every step follows a backup→remediate→verify structure with rollback guidance
- Code-generation agent: screened by prompt-injection input guardrails; outputs must pass a generate-and-review loop with a separate reviewer agent
Tamnoon chose Bedrock to match different models to different jobs via one API (deep reasoning vs. high-throughput validation vs. code review), leverage Cross-Region inference for automatic traffic routing, and keep inference inside IAM-scoped trust boundaries. It integrates with Wiz, CrowdStrike, GuardDuty, and others, normalizing heterogeneous findings into end-to-end remediation.
More from coding & agent
- Consultancy ships free playbook for a product design harness that keeps coding agents on-brand — thisiskp_ · 2026-10-14
- Reward-Rate Policy Gradient: Teaching Agentic RL to Pay for Execution Time — sherryyangML · 2026-10-14
- tldraw plugin runs embeddable canvases inside ChatGPT, with app logic drawn as statecharts — max__drake · 2026-10-14
- MazeBench record: Schema with GPT-6 Astra hits 33%, beating top-50 human median — patience_cave · 2026-10-14
- Perplexity Computer decompiles a Windows-only keyboard app with Ghidra and ports it to native Linux Rust — AravSrinivas · 2026-10-14
- CMU Paper: Most Wrong Actions by System One Models Follow Correct Reading of State — CarnegieMellonU · 2026-10-14