← Founder Notes
Archive

Security researchers found a way to fire agent tools without running the model. amazon bedrock's…

Yethikrishna ROriginal on Threads

security researchers found a way to fire agent tools without running the model. amazon bedrock's agentcore, google's adk and vercel's ai sdk harness all had the flaw, reported september 16.

the tool layer became the attack surface.

Context

The Hacker News, a secondary trade outlet, reported on August 6, 2026 that AWS, Google and Vercel patched agent flaws, citing researchers Hedi Ingber and Aviyam Ivgi, named as co-founders of Stealth, who presented a pattern they call CoreBreak at Black Hat USA 2026. Per that report, an AWS bulletin CVE-2026-18830 (CVSS v4.0 8.6) covered the managed InvokeHarness, fixed before July 31 with no customer action. Google ADK Python before 2.5.0 had CVE-2026-18236 (9.3), a confirmation-forgery path, and Vercel patched @ai-sdk/harness-codex 1.0.29 and @ai-sdk/harness-opencode 1.0.28.

How it compares

No primary vendor advisory was inspected, so versions and scores are unverified beyond that one secondary report. The report says the conditions differ: AWS needed an authenticated remote request, Google needed attacker-controlled session events or user-authored function calls, and Vercel needed untrusted code already inside a Linux sandbox, so all had the flaw compresses three different issues. The researchers say an AWS Strands open-source path remains with documentation and no code fix, which was not independently verified. No source dated September 16 was found, a later Cloud Security Alliance note of September 19 was seen only as a snippet, and no exposure of any particular owner is inferred. The tool layer became the attack surface is the author's take.

Watch next

  • First-party AWS, Google and Vercel advisories.

Sources

  1. The Hacker News: AWS, Google and Vercel patch agent flaws (August 6, 2026)thehackernews.com

Provenance

The note above is reproduced unedited from the original post, first published on Threads on 22 September 2026 at 20:39 IST. Sources are the papers and datasets the note draws on.

View the original post
Embed this note
<iframe src="https://founder.myndlabs.tech/notes/embed/security-researchers-found-a-way-to-fire-agent-DdmBhuUAtpB" width="480" height="420" style="border:0;max-width:100%" loading="lazy" title="Security researchers found a way to fire agent tools without running the model. amazon bedrock's…"></iframe>

More notes