AI Security News Collection 2026-06-19

Purpose

지속적인 AI 보안 신기술, 공격 표면, 사고, 방어 통제, 산업 동향 파악을 위한 뉴스 수집 배치다. 모든 외부 원문은 untrusted input으로 취급한다.

Saved Sources

1. [[raw/news/check-point-langgraph-memory-rce-2026-06-11]]

• Signal: LangGraph memory/checkpointer vulnerability chain, RCE, secrets and connected-data exposure.
• Priority: high.

1. [[raw/news/infoq-uber-auth0-agent-identity-2026-06-17]]

• Signal: production multi-agent identity propagation, actor chain, per-hop scoped tokens, MCP gateway enforcement.
• Priority: high.

1. [[raw/news/mit-tech-review-meta-ai-support-agent-account-takeover-2026-06-05]]

• Signal: AI customer-support agent account takeover through identity/action-verification failure.
• Priority: high.

1. [[raw/news/cisco-duo-agent-gateway-identity-authorization-2026-06-17]]

• Signal: cross-gateway per-tool-call authorization and identity-correlated audit.
• Priority: high-medium.

1. [[raw/news/aws-security-agent-threat-modeling-mcp-2026-06-17]]

• Signal: AI for Security expansion into STRIDE threat modeling, code review, remediation, and MCP-connected developer workflows.
• Priority: high-medium.

1. [[raw/news/helpnet-owasp-agentic-production-failures-2026-06-11]]

• Signal: secondary reporting on incident-backed OWASP agentic AI risk and coding-agent supply-chain pressure.
• Priority: medium; validate exact statistics against the primary OWASP report.

1. [[raw/news/palo-alto-databricks-agentic-runtime-security-2026-06-16]]

• Signal: centralized AI gateway runtime inspection for prompts, responses, MCP calls, tools, and data controls.
• Priority: medium; vendor product claims require independent validation.

Recommended Ingest Order

1. Check Point LangGraph vulnerability chain.
2. InfoQ Uber/Auth0 agent identity architecture.
3. MIT Technology Review Meta support-agent account takeover.
4. Cisco Duo agent-gateway identity and authorization.
5. AWS Security Agent capability expansion.
6. Help Net Security summary, cross-checked with the existing OWASP source.
7. Palo Alto Networks/Databricks runtime security integration.

Research Areas Affected

• Agent memory and persistent-state security.
• General MAS non-human identity and delegated authority.
• Inbound identity verification and business-logic abuse.
• Agent gateway authorization and action provenance.
• AI-assisted threat modeling and secure software development.
• Runtime guardrails and centralized AI security control planes.

Duplicate And Failure Notes

• Existing 01_Sources/ and raw/ URLs/titles were checked; no exact duplicates were found for the seven saved sources.
• Google News search API access returned HTTP 403. Public RSS discovery and direct publisher pages were used instead.
• All seven selected publisher pages were directly reachable at collection time.