Security Analyst | Cloud · Kubernetes · DevOps/SRE · AI-driven threat detection

Polymath approach inspired by Tagore and da Vinci: systems that are both rigorous and elegant.

• Core: Linux, Windows, Bash, C, Java
• Infra: IaC, Docker, Kubernetes
• Observability: Prometheus, Grafana
• Focus: Cloud security (AWS/Azure/GCP), container & K8s security, IaC security, incident response, AI/ML for threat detection

CMS – Cloud Management System
Security-first cloud management with Docker, Kubernetes, Prometheus, Grafana. Monitoring + automated compliance.

![GitHub Stats] (https://github-readme-stats.vercel.app/api?username=hkra1&show_icons=true&theme=radical) ![Top Languages] (https://github-readme-stats.vercel.app/api/top-langs/?username=hkra1&layout=compact&theme=radical)

• Vault Secrets Operator CVE-2026-8715 (CVSS 9.6) — AppRole secretIDPath lets a low-priv tenant force the operator pod to exfiltrate its own ServiceAccount token (cluster-wide Secrets read/write + serviceaccounts/token create). One hop from cluster-admin. Field removed in 1.5.0; treat operators that accept arbitrary paths/URLs as tier-0 attack surface. Write-up · Advisory
• Container security pivots from scan-and-patch to attack-surface reduction — Vulnerability volume + regulatory pressure (NIST, CRA) make pure detection unsustainable. Distro Zero / minimal images (strip OS components the app never needs) remove classes of CVEs before they exist. Scanners still matter; the winning move is fewer things to scan. SiliconANGLE / theCUBE
• Operator-as-confused-deputy pattern keeps repeating — Vault Secrets Operator, External Secrets Operator, Argo CD, Rancher: any controller that runs with elevated privileges and accepts tenant-controlled inputs (paths, URLs, manifests) becomes a privilege-escalation oracle. Default-deny network policies + admission policies that forbid dangerous fields remain non-negotiable.

• llama.cpp critical memory-safety cluster (Cyera / DEF CON) — 10 vulns (UAF, integer overflow, OOB) in the dominant local LLM runtime; two llama-server UAFs at CVSS 9.2 enable race-conditioned unauth RCE when sleep-idle is enabled. Local AI inference is now a first-class network attack surface—auth, isolation, and build pinning are mandatory. eSecurity Planet · VulnCheck
• AI-generated code security pass rate collapses to 56% (Veracode) — 44% of generated samples contain OWASP Top 10 issues; Java fares worst at 30%. Volume without verification is just faster technical debt—treat every AI PR as untrusted until static analysis + human review + runtime proof.
• OpenAI GPT-5.6-Cyber for defenders (Daybreak) — Restricted cyber-permissive model answers 95% of advanced exploit/validation requests (vs ~2% for guarded variants). Capability is real; the access-control and dual-use governance problem is now acute.
• CSA 2026 Top Threats to Cloud Computing — Identity & access management reclaims #1; AI-Enhanced Attacks debut at #2 and AI System Compromise at #6. Traditional infra/CSP concerns drop off the list. Non-human identities, autonomous agents, and supply-chain complexity now outpace governance; treat AI systems as both weapons and high-value assets under zero-trust. CSA Report · Press
• SharePoint CVE-2026-55040 auth bypass actively exploited — Critical JWT validation chain (alg:none + STS cert thumbprint + unverified actor token) lets unauthenticated attackers impersonate any site user/admin after Rapid7 PoC. Fifth SharePoint vuln exploited this year; spike in attempts Aug 12–13. Patch immediately; monitor for forged Bearer tokens. The Hacker News · Rapid7
• VMware vCenter CVE-2026-59310 directory traversal under active exploitation — Network-accessible path traversal → arbitrary code + reverse_ssh persistence. 361 victim IPs across 47 countries observed post-disclosure; outbound reverse SSH bypasses inbound controls. Patch Broadcom appliances; hunt for unexpected reverse_ssh + outbound connections on vCenter hosts. The Hacker News
• Stealing Reasoning Traces from Proprietary LLM APIs — Encrypted CoT blocks from OpenAI, Anthropic, and Google could be replayed into weaker sibling models in the same family, recovering hidden reasoning (and secrets) in plaintext. 315k+ blocks decoded from public agent logs yielded hundreds of credentials/PII; also enables distillation and opaque prompt-injection. Providers mitigated the cross-session/oracle path; treat client-carried “encrypted” state as untrusted. The Hacker News · arXiv
• LiteLLM / TeamPCP AI supply-chain scale (updated) — CloudSEK mapping of the March LiteLLM compromise (via poisoned Trivy in CI) now ties ~2,500 orgs and ~434k CI/CD pipelines to harvested cloud keys, K8s tokens, GitHub secrets, and LLM API keys. 40-minute PyPI window; blast radius is long-lived credentials + AI gateways as privileged CI. Pin by SHA, rotate everything the runner could read, treat AI infra as tier-0. The Hacker News · CloudSEK
• CrowdStrike real-time CDR + K8s control-plane detections for GKE — Falcon Cloud Security extends real-time cloud detection/response to Google Cloud and adds Kubernetes audit-log based threat detection for the control plane (service accounts, secrets, orchestration abuse). Runtime sensors alone miss API-driven lateral movement; correlate control-plane + workload + identity. CrowdStrike
• Windows Container Isolation FS filter (CVE-2026-72971) — August Patch Tuesday includes a tampering vuln in unionfs.sys (Windows container isolation driver). Host-adjacent container surfaces remain attack vectors even on Windows; keep isolation boundaries tight and patch promptly.

• LinkedIn
• GitHub
• YouTube
• Twitch
• Discord: hkra1

Open to: security/DevOps collaboration, SRE roles, mentoring, AI-security research.