Hive Metastore SQLi scores 9.8, PraisonAI auth fails wide open
A SQL injection in Apache Hive partition names (CVE-2026-49845, CVSS 9.8) plus an "auth fails open" bypass in PraisonAI (CVE-2026-55533, CVSS 8.2). Also: Coroot OAuth hijack, LACT Polkit race condition, and NVIDIA NemoClaw command injection.
No fires today, but one worth your attention. CVE-2026-49845 is a SQL injection in Apache Hive Metastore that scores a 9.8. The real blast radius is narrower than that number suggests (scoped to partition metadata, requires direct SQL mode), but if you run Hive with default settings, don't sleep on it. Four more bugs round out the list, including an auth-bypass in PraisonAI that fails wide open when no API key is configured.
Today's CVEs
Sorted by urgencyCVE-2026-49845
NVDAn authenticated user who can call Hive Metastore APIs can inject SQL through crafted partition names. Because the metastore concatenates those names into queries instead of using bind parameters, an attacker can read, modify, truncate, or drop partition metadata for any partition in a table they can reference. The CVSS is 9.8, but real-world impact is narrower than typical SQLi: it's scoped to partition-level operations within tables the caller already has access to, and it requires direct SQL mode to be on (which is the default).
- Affected estate
- Anyone running Apache Hive Metastore before 4.2.1 with direct SQL enabled (the default setting)
- How to check
- Run 'hive --version' or check your Hive package version. Confirm the metastore.try.direct.sql property in hive-site.xml (defaults to true).
- Included because
- CVSS 9.8; SQL injection in a default-enabled code path; authenticated but low-complexity attack
- Action
- Upgrade Hive to 4.2.1. As a short-term workaround, disable direct SQL by setting metastore.try.direct.sql=false.
- Why it matters
- Authenticated callers can manipulate partition metadata, potentially truncating or dropping the wrong partitions and corrupting table state.
- Source
- Apache Hive project advisory
Evidence trail
- NVD: View source
CVE-2026-55533
NVDPraisonAI's authentication middleware silently allows all requests when the expected API key or JWT secret isn't configured. If your Recipe server is exposed to the network, anyone can POST to /v1/recipes/run without credentials, even though you think auth is turned on. This is a classic "auth fails open" bug.
- Affected estate
- Anyone running PraisonAI before version 4.6.58 with a network-exposed Recipe server
- How to check
- Check your installed version with 'pip show praisonai'. Verify whether PRAISONAI_API_KEY and PRAISONAI_JWT_SECRET are set in the server's environment.
- Included because
- unauthenticated; internet-facing; authentication bypass; CVSS 8.2
- Action
- Upgrade to PraisonAI 4.6.58 via pip. Restrict network access to the Recipe server endpoint if you can't upgrade immediately.
- Why it matters
- An unauthenticated attacker can execute arbitrary recipe runs on your server, bypassing what you believe is working authentication.
- Source
- PraisonAI GitHub advisory
Evidence trail
- NVD: View source
CVE-2026-79786
NVDCoroot's MCP OAuth endpoint lets anyone register a new OAuth client with any redirect URI, no authentication required. An attacker registers a client pointing to their own server, sends a crafted authorization link to a signed-in user, and captures the authorization code when the user clicks "approve." That code can be exchanged for an access token, giving the attacker full control of the victim's MCP session.
- Affected estate
- Anyone running Coroot with the MCP OAuth dynamic client registration endpoint exposed
- How to check
- Try sending an unauthenticated POST to the dynamic client registration endpoint with an arbitrary redirect_uri. If it succeeds, you're vulnerable.
- Included because
- unauthenticated; internet-facing; session hijack via OAuth redirect; CVSS 7.1
- Action
- Update Coroot to the fixed version. If you can't update yet, block external access to the dynamic client registration endpoint.
- Why it matters
- Attackers can hijack MCP sessions of any signed-in user through a phishing link, gaining their access level.
- Source
- Coroot project advisory
Evidence trail
- NVD: View source
CVE-2026-75037
NVDLACT (a Linux GPU control tool) uses Polkit authentication based on the caller's PID to verify privileged requests. A local attacker can exploit a race condition in PID recycling to bypass authentication and run privileged LACT operations without proper authorization. This requires local access to the machine.
- Affected estate
- Linux users running LACT version 0.10.0 or earlier for GPU management
- How to check
- Run 'lact --version' or check your package manager for the installed LACT version.
- Included because
- local privilege escalation; authentication bypass; CVSS 7.0
- Action
- Update LACT to the fixed version containing commit d0478fe.
- Why it matters
- A local attacker can bypass Polkit authentication and perform privileged GPU configuration changes without authorization.
- Source
- LACT project advisory
Evidence trail
- NVD: View source
CVE-2026-65099
NVDNVIDIA NemoClaw's CLI on Linux is vulnerable to OS command injection. A local attacker who can supply crafted input to the CLI can execute arbitrary commands with the privileges of the NemoClaw process. This could lead to code execution, data tampering, or full system compromise depending on how NemoClaw is deployed.
- Affected estate
- Linux users running NVIDIA NemoClaw
- How to check
- Check your installed NemoClaw version against NVIDIA's security bulletin for the fixed version number.
- Included because
- command injection; code execution; CVSS 7.8; common vendor
- Action
- Update NemoClaw to the patched release listed in the NVIDIA security advisory.
- Why it matters
- Command injection in the CLI can give an attacker code execution at the privilege level of the NemoClaw process.
- Source
- NVIDIA security advisory
Evidence trail
- NVD: View source
One email, every Wednesday morning.
SubscribeFrom the field notes
From this beat
Read the rest of the field notes โ