Hi there, this is your daily ☕️ Cyberpresso.
In today's Cyberpresso:
🔓 Hackers exploit Microsoft AI package
🇷🇺 Google detailed how COLD RELIC hackers chained US breaches
☕ Researchers disclosed critical Java4Shell RCE flaw in OpenJDK
🔧 Critical Perforce Puppet flaw is under active attack
🐛 Apache Roller flaw lets attackers hijack servers
Plus: 💡 7 strategies & tactics, 🎁 5 more stories you might like, 🧰 6 tools, and 📚 5 papers.
Hackers exploit Microsoft AI package LINK
- Hackers uploaded two booby-trapped versions of Microsoft's Agent Framework package to the Python Package Index yesterday, hiding a credential stealer that ran during normal Python startup and targeted Azure identities, GitHub credentials, and secrets from automated builds.
- Both malicious releases planted a .pth file that Python runs automatically, so any script in the affected environment triggered the stealer, which hunted Azure CLI data, Git stores, and fresh Azure and GitHub Actions tokens.
- Posted under Microsoft's real PyPI publisher, the files were pulled within an hour, but installed copies remain; the fix is to pin Agent Framework to the clean 1.26.0 release, rebuild from a trusted image, and rotate exposed credentials.
🇷🇺 Google detailed how COLD RELIC hackers chained US breaches LINK
- Google and Mandiant detailed how Russian state hackers known as COLD RELIC broke into one US organization's Google Workspace in late 2025, then used the stolen data to compromise a second US organization in 2026.
- The first breach started with a stolen Super Admin account and MFA fatigue, where attackers spammed login prompts until approval, then exported the staff directory, downloaded the entire cloud drive, and deleted security alerts.
- Posing as a first-victim employee over an encrypted chat app, the hackers phished Microsoft Entra ID logins at the second organization; Google saw fresh phishing as recently as this month, with no malware payload found in either break-in.
Researchers disclosed critical Java4Shell RCE flaw in OpenJDK LINK
- Researchers at Calif Research disclosed a critical remote-code flaw in OpenJDK, the free Java runtime, dubbed Java4Shell and tracked as CVE-2026-114938, letting an attacker seize servers with no password needed.
- The memory bug sits in OpenJDK's Base64 decoder on AVX-512 chips, scoring 9.2 on the CVSS scale; a specific input corrupts memory, and researchers broke into the Keycloak login system to prove full control.
- No real-world attacks had been reported as of today, and Oracle shipped patched Java builds yesterday afternoon alongside OpenJDK's advisory; the team found the flaw in about an hour using OpenAI's Codex.
Critical Perforce Puppet flaw is under active attack LINK
- Hackers are actively exploiting a critical flaw in Perforce Puppet, tracked as CVE-2026-115731, that lets them run their own code on vulnerable configuration-management servers and tamper with the systems those servers control.
- The bug, scored 9.4, abuses how Puppet handles report data: a client with a Puppet-signed certificate can send a malicious report that runs on the server, potentially exposing its certificate-signing key to attackers.
- Perforce patched the hole on November 3 in several Puppet Enterprise and Core versions, and CISA added it to its exploited-bugs list yesterday; a proof-of-concept hitting the related OpenVox platform is already public.
Apache Roller flaw lets attackers hijack servers LINK
- Hackers are actively exploiting a critical flaw in Apache Roller, an open-source Java blogging server, letting unauthenticated attackers run their own code on vulnerable systems, Apache warned in a security advisory yesterday.
- The bug, tracked as CVE-2026-129448, sits in Roller's XML-RPC endpoint, which deserializes vendor extension types before checking who sent the request, affecting most releases before the latest patch.
- Apache urges users to upgrade to Roller 6.1.6 or later, which turns off the XML-RPC feature; admins who can't patch yet can disable extensions and remove the servlet mapping in web.xml, then restart.
Strategies & Tactics
> Race Against the (Mac)hine: Local Privilege Escalation in Microsoft Defender for macOS: A race condition in Microsoft Defender for Mac lets a regular user impersonate a trusted process and gain root access, showing how patches don't always carry across operating systems.
> Researchers Break AMD SEV-SNP Security to Read Encrypted Virtual Machine Memory Using a GPU: Researchers bypassed AMD's confidential-computing protection by misdirecting setup writes to a graphics card, letting a malicious cloud host read and alter guest virtual machine memory.
> Cloud keys leaked in Docker Hub images are stolen within seconds: Automated scanners strip API keys from public Docker container images within seconds of upload, so leaked cloud credentials must be considered compromised instantly.
> Depth-Conditioned Backdoors in Looped Language Models: A hidden trigger can be planted in looped AI models that activates only at certain computation depths and survives safety retraining, exposing a new oversight weakness.
> Secure Exclaves and Ambient Sensing: Apple's new Secure Exclaves wall off camera and microphone data in an isolated chip environment so background sensing like "Hey Siri" stays private and shows a tamper-proof indicator when sensors go fully active.
> Inside the loop: Five AI attack frameworks that never ask permission: Attackers now wrap AI planning loops around standard hacking tools, so defenders must base detections on behavior rather than quickly-changing fingerprints or source addresses.
> Is AI pentesting as good as a human pentester?: AI pentesters match or beat humans by reading entire codebases for the price of four tests, enabling frequent deep testing, though they over-rate risk.
Other news you might like
- httpx is vulnerable to Remote Code Execution via Oj Object-Mode JSON Deserialization LINK
- 4-Year-Old NVIDIA Linux GPU Flaw Lets Users Hijack Entire Shared Servers LINK
- TheMailer – Self-hosted Phishing Platform that Runs on Claude and ChatGPT LINK
- BubbleLoader malware turns copies of WinRAR and Windows Media Player into carriers LINK
- Critical Apache HTTP Server RCE Vulnerability Actively Exploited in Attacks LINK
Trending tools
VibeDefend: embeds security and business rules into AI coding agents like Cursor and Copilot, catching vulnerabilities and malicious packages before code reaches your PR. LINK
face2social: reverse image search that matches a face across Instagram, Facebook, LinkedIn, TikTok, and X to verify identities and catch scammers. LINK
Upsolve Row-Level Security: enforces data access permissions server-side, so AI agents can safely serve multiple tenants without bypassing or exposing row-level rules. LINK
Off the Record: emits inaudible interference that blocks AI notetakers from transcribing meetings, including hidden ones running silently on other people's devices. LINK
Koreshield: screens customer messages, documents, and tool calls before your AI agent acts, catching data leaks, hidden instructions, and unsafe actions while logging decisions. LINK
AllModels: a unified OpenAI-compatible API for accessing abliterated models, letting you compare providers, latency, pricing, and capabilities through one endpoint. LINK
Trending papers & reports
Factory-maintenance AI gets a physics-based safety check that blocks tampered sensor data and malicious documents from faking machine health, cutting a hidden attacker's success on healthy equipment from ~53% to ~5% while rarely crying wolf. LINK
Byzantine-robust hospital AI breaks down in real medical collaborations of just four to six hospitals, where defenses meant to block bad actors instead let one data-heavy hospital dominate or cripple the shared model. LINK
Vibe coding security gets its first unified threat map, cataloging twenty-one ways AI coding agents can be exploited and warning that risks grow as agents gain more authority to act on their own. LINK
Swedish government PDFs leak sensitive hidden data in nearly half of cases, with recoverable deleted text including personal details found in 228 files, exposing a persistent gap between new cybersecurity laws and actual publishing habits. LINK
Recommendation-system fraud detection catches sneaky attempts to trick AI-powered product recommenders into pushing certain items, with a watchdog tool that keeps learning new scam patterns and explains its reasoning, protecting next-generation recommendation platforms. LINK
See you tomorrow for a new dose of ☕️ Cyberpresso!