security vulnerability

Bilingual

Coinkite warns Coldcard hardware wallet users of firmware flaw
2 months ago
- Coinkite warns that COLDCARD Mk3 firmware versions 4.0.1 to 5.0.3 may have vulnerable wallet seeds, putting Bitcoin at risk.
- The issue does not affect COLDCARD Mk4, Mk5, or Q devices; users with a BIP-39 passphrase face minimal risk.
- Users are urged to migrate to a new seed generated on unaffected hardware, verifying backups and testing transfers before moving large balances.
- Coinkite is investigating and will release a detailed technical analysis; experienced users may use a dice-roll method to avoid the Mk3's random-number generator.
CosmosEscape: Taking over Every Database in Azure Cosmos DB
2 months ago
- Wiz Research discovered a critical vulnerability in Azure Cosmos DB's Gremlin API, named CosmosEscape, that could allow attackers to compromise all databases in the service, including Microsoft's internal databases.
- The vulnerability enabled attackers to retrieve a platform-wide secret called the Cosmos Master Key, granting full read/write access to any Cosmos DB account and the ability to enumerate all databases by subscription or tenant ID.
- Cosmos DB is used internally by Microsoft services like Microsoft Entra ID, Microsoft Teams, and Microsoft Copilot, making them potentially accessible via this vulnerability.
- Microsoft has fully remediated the issue by eliminating the Cosmos Master Key and introducing new guardrails, with no customer action required.
- The vulnerability was exploited by bypassing the Gremlin sandbox through .NET reflection, allowing arbitrary code execution on the DB Gateway, which then used the master key to access other accounts.
- The attack chain involved querying the Config Store to enumerate accounts and using the master key to retrieve primary keys, affecting both customer and Microsoft internal databases, including network-isolated accounts.
- Microsoft deployed a hot fix within 48 hours and completed a long-term architectural fix by July 2026, with no evidence of exploitation beyond the research.
Document-borne AI worms can self-propagate through Copilot for Word
2 months ago
- Attackers can embed hidden instructions in Word documents that cause Microsoft Copilot to alter content and propagate the attack to new documents, creating a self-replicating AI worm.
- The attack exploits the trust boundary between attached documents and the current drafting document, as Copilot treats all text in its context as potentially authoritative.
- No robust mitigation exists for the broader vulnerability class; Microsoft's fixes closed specific payloads but not the underlying architectural weakness.
- Once propagated, the attack becomes difficult to trace because affected documents are created by legitimate users and may spread across organizations via shared SharePoint or Teams.
- The underlying issue is that LLMs cannot reliably separate attacker-controlled content from trusted instructions, a fundamental architectural challenge in current AI systems.
24,650 internet-accessible BMCs leak password-derived hashes before login
2 months ago
- A 20-year-old vulnerability (CVE-2013-4786) in IPMI 2.0 allows offline password cracking by exposing password-derived hashes before authentication.
- Over 36,000 data center servers exposed IPMI interfaces to the public internet, with 24,650 vulnerable to hash disclosure.
- Weak or factory-default passwords on BMCs (e.g., Supermicro and HPE) are recoverable via GPU-based cracking within minutes to hours.
- Compromised BMCs provide privileged, persistent access below the operating system, enabling lateral movement across management networks.
- AI infrastructure is especially at risk due to shared management networks and high-value GPU systems.
- Recommendations include blocking UDP port 623, replacing default passwords, isolating BMCs on private networks, and continuous monitoring.
StubZero: $148,337 RCE in Google Cloud Production
2 months ago
- Discovered a debugging endpoint on Google Cloud's internal API that leaked protobuf definitions for any service, including YouTube.
- Leaked internal workflow execution queue containing client_id 'default', enabling further access.
- Collaborated with another researcher to bypass publishing restrictions and achieve Remote Code Execution via GenericStubbyTypedTaskV2.
- First RCE was awarded $60,000 by Google's VRP.
- Three months later, found an IDOR vulnerability in the public Application Integration product that leaked test case IDs across projects.
- Used binary search on test case filter parameter to extract victim integration version UUIDs, enabling cross-tenant access.
- Attempted to chain IDOR with internal task types like PythonTask and GenericStubbyTypedTaskV2 for another RCE, confirmed exploitable.
- Second RCE awarded $75,000, plus $13,337 for a remaining vulnerable endpoint.
Leaking the phone number of any Google user
2 months ago
- The username recovery form on Google still works without JavaScript, using two HTTP requests to check if a recovery email or phone number is associated with a display name.
- Brute-forcing phone numbers is possible by using IPv6 address rotation to bypass rate limits, but datacenter IPs still face captchas due to JS disabled form restrictions.
- Using BotGuard tokens from the JS-enabled form with the No-JS form bypasses request limits, allowing efficient brute-forcing of phone numbers.
- Google's display name leak via Looker Studio enables attackers to get the victim's name without interaction.
- The full attack chain involves leaking the display name, obtaining the masked phone from the forgot password flow, and brute-forcing the phone number with optimized tools.
- Brute-forcing times vary by country: 20 minutes for the US, 4 minutes for UK, 15 seconds for Netherlands, and 5 seconds for Singapore, with potential reduction using hints from other services.
Skills That Don't Exist: A Large-Scale Study of Hallucinated Skill
3 months ago
- LLM agents recommend and install skills from open registries, but frequently hallucinate skill names that do not exist.
- Skill name hallucination enables supply-chain attacks, where adversaries pre-register malicious skills under fake names for victims to install.
- A large-scale study evaluated 15,000 prompts across 12 configurations, revealing hallucination rates averaging 36.0% for standalone LLMs and 36.9% for agents.
- Hallucinated names are not random; agents repeat the same fake names, providing reliable targets for hijacking.
- Defenses like retrieval grounding reduce hallucination but severely impact usability, highlighting a security-usability conflict.
- Fixing this vulnerability requires structural ecosystem changes, such as registry-level name reservations and verified recommendation pipelines.
Microsoft's Secure Boot has been broken for a decade and no one noticed
3 months ago
- Microsoft's Secure Boot protection, designed to prevent firmware infections on Windows and Linux devices, has been easily bypassed for most of its existence due to unrevoked, vulnerable shim images.
- ESET researchers discovered 11 old firmware images, or shims, dating back to at least 2013, that remain signed by Microsoft despite known vulnerabilities, allowing attackers to circumvent UEFI Secure Boot.
- Attackers can use these shims to install malicious firmware (bootkits) that persist through OS reinstallation or hard drive replacement, affecting both Windows and Linux users.
- Secure Boot was introduced in 2012 to counter bootkits, but its bypass has been simple, requiring only an old, trusted shim binary and basic UEFI knowledge, undermining this critical security feature.
- Many Linux users disable Secure Boot due to its ineffectiveness on most hardware, citing requirements like proper UEFI defaults and a fully local, auditable chain of trust that are rarely met on consumer-grade devices.

security vulnerability

Coinkite warns Coldcard hardware wallet users of firmware flaw

CosmosEscape: Taking over Every Database in Azure Cosmos DB

Document-borne AI worms can self-propagate through Copilot for Word

24,650 internet-accessible BMCs leak password-derived hashes before login

StubZero: $148,337 RCE in Google Cloud Production

Leaking the phone number of any Google user

Skills That Don't Exist: A Large-Scale Study of Hallucinated Skill

Microsoft's Secure Boot has been broken for a decade and no one noticed