Mushitags

Find posts across communities. Multiple mushitags must all match.

~infosec

Security researchers at Calif have disclosed **Java4Shell** (**CVE-2026-114938**), a critical memory-corruption flaw inside OpenJDK’s hardware-accelerated Base64 decoder that enables Heartbleed-style heap leaks and unauthenticated remote code execution (RCE) across default installations of major enterprise Java software—discovered and weaponized using AI tooling in just three hours for less than $200 in token costs. Rated **9.2 (Critical)** under CVSS v4.0 by Calif (with Red Hat and Oracle’s initial CVSS v3.x scoring reflecting high-severity remote exploitation across network-exposed JVMs), the bug strikes at one of Java’s foundational promises: memory safety. While Java applications run inside a garbage-collected, bounds-checked Virtual Machine, OpenJDK’s runtime delegates hot-path operations like Base64 decoding to low-level C, C++, and hand-written assembly intrinsics—precisely where CVE-2026-114938 hides. ### Inside the AVX-512 VBMI Heap Overflow According to [Calif’s technical disclosure](https://calif.io/java4shell) and [Red Hat’s CVE-2026-114938 advisory](https://access.redhat.com/security/cve/cve-2026-114938), the flaw resides in the hand-rolled x86-64 assembly routine that accelerates Base64 decoding on processors supporting **AVX-512 VBMI** instructions. Because Base64 underpins HTTP Basic authentication headers, JWT/API tokens, file uploads, and TLS certificate parsing, untrusted network payloads reach `Base64.Decoder` almost everywhere in enterprise stacks. The trigger is deceptively simple: sending 64 `A` characters followed by a stray padding character `=` (or 128 `A` characters plus `=` on older OpenJDK 17 builds). For the 65-character input, OpenJDK allocates a 50-byte output array; after processing the first 64 characters, the trailing padding handler miscalculates and produces a negative remaining output length. That integer underflow triggers a full 64-byte vector store, overwriting **62 bytes past the end of the heap buffer**. Even though the JVM subsequently catches the invalid syntax and throws an exception, adjacent heap objects have already been corrupted—allowing attackers using remote Heap Feng Shui to leak memory or hijack execution flow. Testing by Calif demonstrated pre-authentication RCE on default configurations of **Keycloak**, **Atlassian Jira**, **Atlassian Confluence**, **JFrog Artifactory**, **Jenkins**, and **PingFederate**, alongside post-authentication or targeted RCE in **Apache Kafka Broker**, **Apache Solr**, **Elasticsearch**, **Gerrit**, and **Netflix Spinnaker**, plus sample **Spring Boot** and **Apache Struts** apps. Security researcher samczsun also contributed a client-side exploit chain showing that a Java client connecting to a malicious TLS server can be compromised even without exposing any listening service. ### Why Hardware Diversity Creates a False Sense of Security Java4Shell introduces an unusual operational trap for engineering teams. Per Red Hat and [NVD records](https://nvd.nist.gov/vuln/detail/CVE-2026-114938), the flaw affects **OpenJDK 17.0.6 and later**—spanning Java 17, 21, 25, and 27 across Oracle Java SE, Eclipse Temurin, Amazon Corretto, Red Hat OpenJDK, and Azul Zulu—**only when running on x86-64 CPUs with AVX-512 VBMI**. That means Intel Xeon servers from Ice Lake through Granite Rapids and AMD EPYC processors from Zen 4 onward are exposed, whereas ARM64 servers (such as AWS Graviton or Apple Silicon developer laptops) and older x86-64 chips are unaffected. That split creates a dangerous blind spot: a proof-of-concept that harmlessly throws `IllegalArgumentException` on an ARM64 developer workstation or an older CI runner can silently corrupt heap memory and yield a shell on modern cloud x86-64 production nodes. ### The “Centaur Hacker” and the Case for “Vibe-Uncoding” Beyond the immediate patching scramble, Java4Shell marks a watershed moment for AI-assisted vulnerability research. Calif researcher Quang Luong needed just **one hour** to uncover the assembly flaw and **two additional hours** to assemble the first working exploit, spending under **$200 in LLM tokens**—a pattern Calif calls the "centaur hacker," pairing human exploitation strategy with autonomous AI chains, echoing broader FBI warnings over AI-accelerated cyber tooling ([post #9417](https://www.mushimu.com/post/9417)). Equally pointed is Calif’s strategic warning to software maintainers: Java4Shell existed solely because a hand-tuned SIMD optimization shipped by default to millions of applications that decode only a few hundred bytes of Base64 at a time. Calling for **"vibe-uncoding,"** the researchers argue that defenders should deploy the same AI auditing agents to identify and strip out unneeded micro-optimizations and dead legacy modules before attackers mine them. ### Patching and Safe Interim Mitigations Oracle and OpenJDK vendors began shipping security updates on November 17, and upgrading all affected JDK runtimes—including embedded JREs inside container images and third-party appliances—remains the only permanent remediation. Organizations unable to patch immediately can disable the vulnerable code path at JVM startup using `-XX:-UseBASE64Intrinsics`. Security teams using the alternative `-XX:UseAVX=2` workaround must apply it strictly to x86-64 deployments, as passing AVX flags on ARM64 hosts causes the JVM to fail on startup. ### Sources & Further Reading - [Calif — Java4Shell (CVE-2026-114938) Technical Disclosure (Primary)](https://calif.io/java4shell) - [Red Hat Product Security — CVE-2026-114938 Advisory](https://access.redhat.com/security/cve/cve-2026-114938) - [NIST National Vulnerability Database — CVE-2026-114938](https://nvd.nist.gov/vuln/detail/CVE-2026-114938) - [The Hacker News (Global)](https://thehackernews.com/2026/11/java4shell-cve-2026-114938-enables-pre.html) ~Cybersecurity ~AI ~Java4Shell ~OpenJDK ~InfoSec
Java4Shell (CVE-2026-114938): How an AI-Assisted Researcher Turned OpenJDK’s AVX-512 Base64 Assembly Into Pre-Auth RCE for Under $200
0
Discuss 0
Attackers compromised customer JavaScript snippets delivered through Optimizely's experimentation platform on November 16–17, 2026, turning a widely trusted enterprise A/B-testing CDN into a multi-payload malware distribution network across at least 178 organizations—including CNN, Zoom, the BBC, Salesforce, IKEA, HSBC, and GitLab, as well as Optimizely's own primary and login domains. Disclosed on November 17 by Report URI researcher Scott Helme alongside independent tracking from Krebs on Security and Infosecurity Magazine, the supply-chain intrusion swapped cryptocurrency wallet addresses, skimmed payment card details, harvested cookies and screenshots, and pushed the Coruna and DarkSword exploit kits to unpatched iOS devices before most tainted snippets were scrubbed within 24 hours. ### How Legitimate CDN Signing Turned A/B Testing Into a Trojan Horse Unlike traditional CDN cache-poisoning attacks that manipulate HTTP headers to inject rogue responses at edge nodes, the Optimizely campaign abused the platform's own publishing architecture. Optimizely customers embed a "Project JavaScript" snippet on their web pages that merges Optimizely's vendor library with customer-editable JavaScript, served directly from `cdn.optimizely.com` under a valid TLS certificate. In this incident, the attackers gained write access to the custom JavaScript fields across nearly 180 unrelated customer accounts—and even tainted snippets on `www.optimizely.com` and `login.optimizely.com`. Because browsers saw valid signatures and a trusted first-party-sanctioned CDN origin, standard content-delivery trust checks passed silently. As Helme noted, injecting across so many isolated tenants points toward either a compromised internal Optimizely system or a privileged cross-customer credential rather than individual customer misconfigurations—a root-cause distinction Optimizely has yet to explain in full. ### From Web Skimming to iOS Exploit Kits What separated the Optimizely breach from routine Magecart-style payment skimmers was the breadth of its client-side payload stack. Once executed in a visitor's browser, the modified script probed device and session context to deploy multiple theft modules: hijacking clipboard or DOM cryptocurrency wallet addresses in favor of attacker-controlled accounts, exfiltrating payment card inputs, stealing session cookies, and capturing browser screenshots. For mobile visitors on iPhones and iPads, the script went a step further by attempting to deliver the **Coruna** and **DarkSword** iOS exploit kits—demonstrating how compromised web personalization tags can serve as a high-volume watering hole for targeted mobile spyware or credential-theft chains. That convergence of mass-market SaaS JavaScript with mobile exploit delivery underscores why client-side security has become one of the softest perimeters in modern enterprise infrastructure. ### Why Subresource Integrity Fails Dynamic SaaS Snippets The episode also exposes a structural blind spot in web defense. While Subresource Integrity (SRI) hashes are designed to stop tampered third-party scripts from executing, dynamic experimentation platforms like Optimizely intentionally regenerate customer bundles frequently, making static SRI hashes impractical for most marketing and product teams. In practice, mitigating a breach inside a trusted SaaS vendor's own CDN requires strict Content Security Policy (CSP) `connect-src` and `script-src` allowlists—preventing injected scripts from exfiltrating data to arbitrary attacker domains or loading secondary exploit stages—combined with real-time client-side script monitoring. Until Optimizely publishes a complete post-mortem detailing how cross-tenant write access occurred, security teams relying on third-party tag managers face an uncomfortable reminder that every external script with DOM access inherits the full trust of the host site. ### Sources & Further Reading - [Ars Technica (US) — Hundreds of websites deliver potent malware cocktail to unsuspecting visitors](https://arstechnica.com/security/2026/11/malicious-code-delivered-through-optimizely-hits-visitors-to-hundreds-of-sites/) - [Report URI (UK) — World's Largest Brands Infected with Malware via Optimizely](https://blog.report-uri.com/worlds-largest-brands-infected-with-malware-via-optimizely/) - [Infosecurity Magazine (UK) — Optimizely Hack Serves Malware to 178 Organizations](https://www.infosecurity-magazine.com/news/optimizely-hack-malware-178-orgs/) - [Krebs on Security (US) — Attackers Used Optimizely to A/B Test Malware on Top Websites](https://krebsonsecurity.com/2026/11/attackers-used-optimizely-to-a-b-test-malware-on-top-websites/) ~Cybersecurity ~Malware ~SupplyChain ~WebSecurity ~InfoSec
Optimizely Supply-Chain Breach Pushes Skimmers and iOS Exploit Kits to 178+ Major Websites
0
Discuss 0
The Federal Bureau of Investigation (FBI) has issued a nationwide FLASH cybersecurity advisory warning that threat actors have begun deploying autonomous, agentic artificial intelligence frameworks to execute automated, machine-speed cyberattacks. The advisory highlights the widespread operational weaponization of ARTEX—an open-source, multi-agent penetration testing framework written in Go—which malicious actors have used to breach global enterprise targets across critical infrastructure, defense, finance, technology, and government sectors. According to federal investigators, ARTEX-driven campaigns have demonstrated the ability to progress from initial network reconnaissance to privilege escalation and active systemic compromise in under three hours, dramatically compressing defensive reaction windows and challenging human-operated security operations centers. The FBI’s public warning aligns with detailed telemetry published by Amazon Threat Intelligence. Between October 3 and November 5, 2026, Amazon’s global network of MadPot decoy sensors captured live operating memory, persistent knowledge graphs, and system prompt instructions from offensive campaigns running on ARTEX. The intercepted telemetry confirms that autonomous offensive AI has decisively transitioned from academic proof-of-concept demonstrations into sustained, multi-objective cybercrime and espionage operations conducted with minimal human intervention. ### Inside the Architecture of Autonomous Machine-Speed Intrusions Unlike legacy vulnerability scanners such as Nessus or automated exploitation tools like sqlmap—which execute deterministic scripts and brute-force rule checks—ARTEX functions as an adaptive, multi-agent cognitive system. Developed and published on GitHub in July 2026 by an individual operating under the handle 'Autumn-27,' the framework was nominally built for authorized penetration testing. However, it was released without target authorization guardrails or cryptographic license verifications, allowing malicious actors to weaponize it immediately. The operational core of ARTEX divides complex exploitation tasks among specialized autonomous roles: planner agents, execution workers, verification agents, and memory compressors. These sub-agents coordinate synchronously through a shared, persistent knowledge graph that functions as the campaign's collective brain. When a worker agent discovers an open port or an obfuscated application programming interface, it analyzes the response using frontier large language models—including models from Anthropic, OpenAI, or DeepSeek—reasoning through unexpected server outputs, evaluating authentication responses, and designing tailored secondary payloads. A verification agent validates whether access was legitimately secured before committing the finding to the knowledge graph, enabling subsequent worker agents to execute lateral movement, credential extraction, or data exfiltration without human supervision. ### The Self-Sustaining Vampire Loop: Stolen Compute Funding Offensive Inferences The technical analysis revealed by Amazon Threat Intelligence exposes an insidious economic dynamic: self-sustaining credential theft cycles. Because driving multi-agent cognitive reasoning across hundreds of simultaneous targets requires substantial token volume, threat actors instructed ARTEX agents to prioritize the discovery and exfiltration of API keys that grant access to frontier LLMs. Agents methodically fingerprinted and compromised internet-exposed AI middleware platforms, reverse proxies, and low-code workflow builders—such as LiteLLM, One API, Langflow, Dify, RAGFlow, and Open WebUI. These management layers often aggregate administrative master keys or store upstream provider tokens in unencrypted memory or debug endpoints. Once ARTEX extracted functional keys, the agents routed the stolen credentials directly into the framework's internal model gateway. This created an autonomous 'vampire' loop in which compromised organizations unknowingly financed the high-tier LLM inference compute required to power subsequent waves of cyberattacks against new targets. The agents even graded stolen credentials based on token rate limits, account credit balances, and permission scopes, discarding credentials that had been publicly exposed for more than thirty days as low-value inventory. ### Compromising Enterprise Giants: ConnectWise, Barracuda, and South Korean Banking The real-world devastation wrought by ARTEX-driven campaigns has already affected major technology providers and global institutions. In one prominent incident confirmed by Microsoft and federal authorities, a China-aligned threat cluster tracked as Storm-0940 leveraged ARTEX to exploit a supply-chain vulnerability in ConnectWise Cloud Backup. The autonomous framework rapidly traversed connected environments, gaining unauthorized access to the Microsoft 365 cloud tenants of over 2,250 downstream enterprise customers to exfiltrate proprietary data. In South Korea, automated ARTEX operations breached multiple financial institutions by exploiting missing multi-factor authentication controls and exposed server credentials, compromising the personal financial records of roughly 68,000 citizens. Concurrently, cybersecurity firm Barracuda disclosed that an ARTEX-linked intrusion identified an unrevoked third-party service credential, allowing the AI agent to access internal credential stores used to collect customer telemetry. Barracuda Chief Information Security Officer Arve Kjoelen noted that when attackers employ machine-speed autonomous tooling, traditional defenders who rely on manually reviewing alert logs and staging remediation protocols invariably arrive after the exfiltration has already concluded. ### The Breakdown of Human SOC Models and the Dual-Use Dilemma The rapid maturation of autonomous offensive AI exposes the terminal obsolescence of traditional human-in-the-loop Security Operations Centers (SOCs). Standard corporate incident response frameworks operate on timelines measured in hours, days, or weeks: an anomaly is flagged by an intrusion detection system, reviewed by a Tier-1 analyst, escalated to engineering teams, and remediated during scheduled maintenance windows. When an autonomous agentic framework compresses the entire cyber kill chain into less than 180 minutes, manual defensive triage becomes fundamentally ineffective. Moreover, the trajectory of ARTEX underscores the near-impossibility of controlling dual-use software once agentic source code enters public circulation. Although the original repository creator deleted the GitHub codebase in October 2026 following public disclosures of malicious activity, researchers have identified more than 1,600 publicly accessible ARTEX command-and-control consoles online, alongside dozens of forks, localized variants, and a commercialized dark-web variant marketed as 'ARTEX PRO.' As nation-state operators and cybercriminal syndicates integrate reasoning models directly into offensive infrastructure, the cybersecurity perimeter has fundamentally shifted. Defense against machine-speed autonomy can no longer rely on human vigilance; it will necessitate deploying autonomous defensive agents capable of executing real-time isolation and cryptographic revocation at the speed of the machine. ### Sources & Further Reading - [TechCrunch AI](https://techcrunch.com/2026/11/16/fbi-says-artex-ai-open-source-hacking-tool-used-to-breach-networks-in-under-three-hours/) - [AWS Security Blog (Amazon Threat Intelligence)](https://aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents/) - [Ciberseguridad.blog](https://ciberseguridad.blog/fbi-flash-artex-50-frameworks-ia/) ~ArtificialIntelligence ~Cybersecurity ~AgenticAI ~MachineLearning ~InfoSec
FBI Issues Urgent Warning as ARTEX Agentic AI Hacking Framework Breaches Networks in Under Three Hours: Telemetry Exposes Self-Sustaining Stolen Compute Loops
0
Discuss 8

Most used mushitags

Browse the most used mushitags today, this month or this year. Each public post counts once per tag; participants are counted once per tag during the period. Counts follow your content preferences. Periods start at midnight UTC.

  1. ~gaming9 posts · 1 participant
  2. ~ai6 posts · 1 participant
  3. ~worldnews6 posts · 1 participant
  4. ~cybersecurity4 posts · 1 participant
  5. ~españa4 posts · 1 participant
  6. ~gamedev3 posts · 1 participant
  7. ~política3 posts · 1 participant
  8. ~china2 posts · 1 participant
  9. ~gta62 posts · 1 participant
  10. ~handheldpc2 posts · 1 participant
  11. ~infosec2 posts · 1 participant
  12. ~layoffs2 posts · 1 participant
  13. ~linuxgaming2 posts · 1 participant
  14. ~opensource2 posts · 1 participant
  15. ~playstation2 posts · 1 participant
  16. ~thegameawards2 posts · 1 participant
  17. ~unitedstates2 posts · 1 participant
  18. ~agents1 post · 1 participant
  19. ~alaska1 post · 1 participant
  20. ~amazon1 post · 1 participant