Amazon Kiro "Power Leak": Kiro Powers prompt-injection data exfiltration
Summary
Mindgard (Fergal Glynn) published "Power Leak: Amazon Kiro IDE Prompt Injection Enables Data Exfiltration" on August 27, 2026. An attacker-controlled repository or page context can steer the Amazon Kiro agentic IDE into rewriting its own MCP server configuration and transmitting sensitive local workspace data to an external endpoint — without the user explicitly asking Kiro to read or send anything. Mindgard assessed exploitation difficulty as low for the tested configuration (Kiro IDE 0.7.45 on Windows; latest public version at publication: 1.0.337). No CVE was assigned; Amazon fixed the flaw in Kiro IDE 0.8.140.
The durable value is twofold: (1) a concrete AI-IDE trust-boundary pattern — steering-file / MCP-configuration writes as an exfiltration and code-execution primitive — and (2) a public case study of disclosure-process breakdown against AI agent products: the related steering-file finding was classified a "duplicate" on HackerOne in December 2025 and sat unremediated until late January 2026, while the Kiro Powers exfiltration path took months to reach publication.
Tags
- patterns
- Amazon Kiro
- Kiro Powers
- prompt injection
- indirect prompt injection
- MCP
- MCP configuration
- steering file
- POWER.md
- IDE trust boundary
- AI IDE
- agentic IDE
- data exfiltration
- credential theft
- workspace trust
- vulnerability disclosure
- HackerOne
- Mindgard
- developer tooling
- AI developer tooling
The mechanism
- Kiro Powers bundle MCP server configurations, steering files (
POWER.md), hooks, and contextual knowledge; the steering file acts as a persistent "onboarding manual" that tells the agent which MCP tools exist and when to use them. - The flaw: attacker-controlled repository content can influence the Kiro agent enough that it rewrites its own MCP server configuration file and causes sensitive local information to be transmitted to an external endpoint — no approval prompt is shown to the user; the developer only asked Kiro to perform a legitimate action.
- Mindgard's write-up frames it as an extension of the earlier steering-file directive finding, where steering directives caused local information to be folded into a Markdown image request delivered to an attacker server (that path published January 15, 2026).
- Precedent in the same product: in June 2026 Amazon fixed an insufficient access-control flaw (CVE-2026-10591, CVSS 8.8) that let a remote unauthenticated actor execute arbitrary commands via crafted instructions writing to execution-sensitive paths such as
.vscode/tasks.jsonand~/.kiro/settings/mcp.json, with auto-execution on folder open.
Disclosure timeline (Mindgard's account)
- December 6, 2025 — initial Kiro steering-file data-exfiltration vulnerability discovered.
- December 8, 2025 — disclosed to Amazon; classified a duplicate through HackerOne on December 10.
- December 11, 2025 — the Kiro Powers prompt-injection exfiltration path discovered and submitted; HackerOne verified and forwarded to Amazon on December 13.
- January 15, 2026 — initial steering-file disclosure published.
- January 26, 2026 — Amazon verified the Kiro Powers disclosure; remediation released the same day.
- August 27, 2026 — this write-up published.
Durable defender guidance
- Treat agent-config files as execution surfaces:
~/.kiro/settings/mcp.json, workspacemcp.json, steering/POWER.mdfiles, andtasks.json-style auto-exec paths are writable state that an agent can change on an attacker's behalf. Inventory and alert on agent-modified MCP/steering configuration in developer environments. - Upgrade Kiro IDE to 0.8.140 or later; verify any pinned or packaged Kiro builds in CI/developer images.
- Hunt for egress from IDE processes to unusual external endpoints (Markdown image URL patterns,
image-setproxies, new HTTP origins) where the user did not request a fetch — the exfiltration rides on ordinary-looking agent network activity. - For AI IDEs generally: restrict what the agent may write (allow-list config paths), disable auto-execution on folder open where the workflow allows, and treat any repository/page content steering an agent toward new network destinations as a malicious-input event, not a model quirk.
- Disclosure-process angle: for red-teamers, "duplicate" classifications on AI-agent findings deserve re-testing, because interaction-surface bugs (model interpretation × tool access × config writes) rarely fit traditional defect dedup.
Confidence and caveats
- No CVE; fixed-version status is per Mindgard's reporting of Amazon's fix (0.8.140) — verify against Kiro's own release notes before treating 0.8.140 as the official first fixed build.
- Version scoping (0.7.45 tested, 1.0.337 latest at publication) is a Mindgard/HN detail; treat affected-version ranges as approximate.
- The "duplicate" HackerOne classification is Mindgard's account of their report handling; Amazon's internal disposition was not published.
Related pages
- Cursor Windows workspace-path binary hijack (Mindgard's earlier Cursor finding — repo-root
git.exeexecution) - Developer-tool config auto-execution
- Agentic workflow trust-boundary failures
- AI coding-agent symlink write confusion
Sources
- Mindgard: Power Leak: Amazon Kiro IDE Prompt Injection Enables Data Exfiltration — August 27, 2026 (Fergal Glynn)
- The Hacker News: Amazon Kiro Prompt Injection Can Exfiltrate Sensitive Data Through Kiro Powers — August 27, 2026