Security Report · raw.githubusercontent.com

Toc Consulting’s 2026 AWS Security Report Examines IAM and Agentic AI Risks

Oct 21, 2026 · Original Report (PDF) ↗

Report discovered via Awesome Annual Security Reports.

Summary

Toc Consulting’s 2026 AWS security whitepaper examines identity misconfigurations, S3 ransomware, and agentic AI risks, with practical controls, audit checklists, and a 90-day hardening roadmap.

Key points

The whitepaper prioritizes MFA, temporary credentials, least-privilege roles, and organizational guardrails to reduce risks from stolen credentials and excessive permissions.
It describes Codefinger operators using compromised AWS credentials to re-encrypt S3 objects with attacker-controlled SSE-C keys, underscoring the need for protected, independent backups.
Existing AmazonGuardDutyFullAccess v1 policy attachments require manual migration to v2, which restricts delegated-administrator registration to specified GuardDuty services.
For monitoring and response, the guide recommends centralized, tamper-protected CloudTrail logs, selective data-event logging, and EventBridge-triggered remediation of high-confidence findings.
The report explains how prompt injection can turn an AI agent into a confused deputy, exercising permissions broader than those of the requesting user.
It distinguishes Bedrock Guardrails’ content filtering from AgentCore’s external Cedar-based tool-call authorization; AgentCore Policy Controls remained in preview as of February 2026.
Its action plan sequences 30-day baseline improvements before broader 90-day hardening, with agent-specific incident playbooks, MCP tool validation, and human approval for high-risk actions.

Article Details

Publisher: Toc Consulting
Report Period: 2026 edition, published February 2026; primarily discusses 2025–2026 developments
Scope: AWS security architecture, identity, data protection, networking, detection, agentic AI security, and compliance.
Key Statistics: The cited Verizon 2025 DBIR examined 22,052 incidents and 12,195 confirmed breaches; compromised credentials were the initial access vector in 22% of breaches.
According to the cited Verizon 2025 DBIR, third-party involvement in breaches increased from 15% to 30% year over year.
The cited IBM 2025 report found that 13% of organizations experienced breaches of AI models or applications; 97% of those organizations lacked proper AI access controls.
According to the cited Akto 2025 report, only 21% of enterprises had full visibility into agent actions, MCP tool invocations, or data access.
The cited Cisco State of AI Security 2026 report found that 83% of organizations planned to deploy agentic AI capabilities, while 29% felt ready to do so securely.
Recommendations: Replace long-lived human access credentials with temporary role-based access, enforce MFA, delete root access keys, and review unused or overly permissive permissions.
Block unnecessary SSE-C encryption and public S3 access; protect critical data with versioning, immutable storage, and independent cross-region backups.
Enable centralized, multi-region CloudTrail logging with integrity validation, sensitive-resource data events, threat detection, and automated remediation for high-confidence findings.
Constrain AI agents with unique identities, least-privilege execution roles, external action policies, content guardrails, and human approval for high-risk operations.
Pin and verify MCP tool versions, treat tool responses as untrusted data, enable model invocation logging, and test agent containment procedures quarterly.