CVE-2026-129875: REBUILD 4.5.0–4.5.4 password hash disclosure via MCP QueryRecords

A high-severity information disclosure flaw in REBUILD 4.5.0 through 4.5.4 allows authenticated users with low privileges to recover the administrator password hash character by character through the MCP QueryRecords tool.

Christopher Ort
By Christopher Ort · 3 min read
Share
A security researcher at a laptop in a dimly lit office, examining lines of code on a monitor, with a notepad showing a sequence of hexadecimal characters.
Image: AI-generated, i10X
On this page
  1. Quick Take
  2. In plain English
  3. What happened
  4. What you should do
  5. What it means
  6. Sources
i10X Quick Take

Quick Take

  1. What: CVE-2026-129875 is an information disclosure vulnerability in REBUILD 4.5.0 through 4.5.4 that allows low-privileged authenticated users to extract the administrator password hash.
  2. Who is affected: Organizations running REBUILD versions 4.5.0, 4.5.1, 4.5.2, 4.5.3, or 4.5.4.
  3. Attack method: Attackers send LK2 prefix-match filters to the /gw/mcp/sse gateway via the MCP QueryRecords tool, using the returned total count as an oracle to extract the SHA-256 hash one character at a time.
  4. Impact: Once the hash is recovered, attackers can log in as administrator.
  5. What to do: No fixed version is stated in the public sources. Monitor the REBUILD GitHub repository for patches.

In plain English

REBUILD is an open-source business application platform with a built-in AI agent feature. In versions 4.5.0 through 4.5.4, any user with a basic account could ask that feature a series of yes-or-no questions about the administrator’s password. Answer by answer, the attacker could rebuild the scrambled password and then log in as the administrator.

What you should do: If you run REBUILD, limit who has accounts and block the AI agent gateway until a fix is out. If you do not use REBUILD, you do not need to do anything.

What happened

REBUILD versions 4.5.0 through 4.5.4 expose an information disclosure vulnerability tracked as CVE-2026-129875. The flaw resides in the MCP QueryRecords tool, which is part of the platform’s AI agent integration. According to the vulnerability report published at vul.yulun.ac.cn, an authenticated user with low privileges can filter the User.password field by sending LK2 prefix-match queries to the /gw/mcp/sse gateway.

The QueryRecords.java code at commit 741e3e5 shows that the tool accepts filter parameters and passes them to the query engine. The AdvFilterParser.java code at the same commit demonstrates that LK2 filters perform prefix matching on fields. Because the tool does not restrict which fields can be queried, an attacker can target the password field.

The attacker sends a series of queries, each testing whether the password hash begins with a specific character (for example, “a”, “b”, “c”). The system returns a total count of matching records. If the count is greater than zero, the attacker knows the first character. The attacker then appends the next character and repeats the process, extracting the SHA-256 hash one character at a time. Once the full hash is recovered, the attacker can use it to log in as the administrator.

The vulnerability was credited to Wu Yulun and disclosed on November 6, 2026. The NVD entry was published on November 16, 2026, with a CVSS score of 8.8 (high severity). VulnCheck assigned a CVSS v4 score of 8.7.

What you should do

The public sources do not state a fixed version. Organizations running REBUILD 4.5.0 through 4.5.4 should monitor the project’s GitHub repository (getrebuild/rebuild) for security patches. Until a patch is available, consider restricting access to the /gw/mcp/sse endpoint or disabling the MCP QueryRecords tool if it is not required for operations.

What it means

CVE-2026-129875 highlights the risk of exposing sensitive database fields through AI agent tools. The MCP (Model Context Protocol) QueryRecords feature is designed to let AI agents retrieve data from the application, but the lack of field-level access control allowed low-privileged users to query the administrator password hash. The use of a prefix-match filter as an oracle is a classic side-channel attack: the attacker does not see the hash directly but infers it from the system’s responses.

The vulnerability is particularly dangerous because it requires only a low-privilege account and can be exploited remotely over the network. Once the administrator hash is recovered, the attacker gains full control of the REBUILD instance. The flaw underscores the importance of validating which fields AI tools can access and ensuring that sensitive data such as password hashes are never exposed through query interfaces, even indirectly.

REBUILD is not the first case of an MCP tool letting users reach data they should not see. Compare the DQL injection in the Dynatrace MCP Server and the SQL injection in the pgEdge Postgres MCP Server. New cases are tracked in our AI Security hub.

Sources

  • NVD CVE-2026-129875
  • REBUILD QueryRecords.java source code
  • REBUILD AdvFilterParser.java source code
  • Vulnerability report at vul.yulun.ac.cn
  • VulnCheck advisory
Christopher Ort

Christopher Ort

Christopher Ort handles SEO, AI agent workflows and PR & partnerships at i10x.ai, the all-in-one AI workspace built by i10X Global Pte. Ltd. in Singapore. He is a graduate of Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU). He is the author of “The Style Penalty”, a study published on SSRN that analyses 1,576 evaluation decisions by four frontier AI models to show how AI-written CVs fare in automated hiring screens; the research has been covered by heise online and t3n.

Editor's picks

  1. CVE-2026-129877: DeepAudit exposes LLM API keys and Git tokens to unauthenticated attackers
  2. CVE-2026-129875: REBUILD 4.5.0–4.5.4 password hash disclosure via MCP QueryRecords
  3. South Korea Weighs Nuclear Power to Meet AI Electricity Demand
  4. UK and US Partner to Accelerate Nuclear Fusion with AI
  5. Texas and New Mexico Data Center Emissions Rise With AI Buildout