Appsec adventures

On August 26, 2026, the OWASP Open Automated Security Initiative for Software (OASIS) launched as an official OWASP community project. This global initiative combines AI-powered fix automation with human expert validation to remediate open source vulnerabilities at scale. OASIS addresses the critical bottleneck of remediation by generating candidate fixes through automated pipelines, having AppSec professionals validate them for correctness and safety, and then submitting vetted patches to open source maintainers. The project has attracted hundreds of security professionals and founding sponsors AppSecAI, Intigriti, and DryRun Security. Designed to complement other industry initiatives like OpenAI's Patch the Planet and the Linux Foundation's Akrites, OASIS offers a vendor-neutral, community-driven approach that empowers application security professionals to collectively defend against AI-driven exploits and "vibe hacking" threats. The initiative invites participants ...

OpenAI has released GPT-6 Astra, described as its most intelligent and aligned model to date, with strong capabilities in computer use, software engineering, science, and cybersecurity. The model demonstrated the ability to identify and develop zero-day exploits, scoring 100% on ExploitBench and discovering two previously unknown vulnerabilities during internal evaluations. In response to these advanced cyber capabilities and lessons from the Hugging Face incident, OpenAI paused some training to implement stricter alignment protocols and safety processes. Astra now includes human oversight features that can slow, pause, or stop its work and request user review before proceeding with certain actions. The model excels in benchmarks, including 98% on FrontierMath Tier 4 and 88% on SRE-Bench for reverse-engineering tasks. Despite its power, Astra is more likely to refuse advanced cybersecurity tasks such as creating proofs-of-concept. It is rolling out to limited organizations initially an...

On September 2, 2026, Google unveiled Mantis, an open-source framework that uses AI to automate the entire software vulnerability lifecycle—from discovery and candidate triage to reproduction and remediation. The framework addresses the challenge of low-precision AI code scanning, where true positive rates can fall below 7%, by employing multiple AI agents for layered verification and reproducing vulnerabilities in isolated sandbox environments before generating remediation code. Mantis uses hierarchical security summarization to reduce token overhead by over 85% while preserving critical structural information, enabling large-scale repository analysis. It is accessible via Gemini CLI and Antigravity CLI, and includes "mantis-advise" for development-time guidance based on historical vulnerability data. Concurrently, Google announced "Gemini 3.8 Flash Cyber," a security-specialized AI model that improved patch accuracy by 2.6x within the Chrome team and discovered a ...

This August 2026 draft guide from NIST provides practical, actionable methods for using AI to help organizations analyze, plan, implement, and monitor progress toward achieving CSF 2.0 outcomes. It offers structured AI prompts as tools for practitioners to create CSF-related artifacts and illustrates current practices in prompt engineering for CSF implementation. The document contains three notional use cases with example prompts, simulated organizational files for a fictitious company, and tips for getting started. Use Case 1 demonstrates an AI-assisted review of cybersecurity policy and governance alignment with CSF outcomes. Use Case 2 shows how to produce a draft Organization Current State Profile by mapping artifacts and interview notes to CSF outcomes. Use Case 3 illustrates creating a draft CSF target state profile based on internal and industry references. The guide includes specific precautions denoted with /!\ notation where needed, though it does not focus on AI best practic...

This article provides a practical guide to selecting hardened container base images as a foundational security layer for Kubernetes environments, emphasizing that risk is inherited from the base image regardless of other controls like admission policies or runtime monitoring. It evaluates eight leading options based on criteria including minimal packages, active patching, SBOM visibility, provenance, compatibility, and developer usability. The list features Echo (prioritizing CVE-free images with application compatibility), Docker Hardened Images (familiar, secure-by-default), Chainguard Images (minimal Wolfi-based OCI images), Google Distroless (runtime-only without OS extras), Red Hat UBI Micro (enterprise RHEL-based minimal foundation), Ubuntu Chiseled Containers (distroless-style Ubuntu runtime), Iron Bank (hardened images for regulated/government environments), and Bitnami Secure Images (hardened application images for common open-source components). The article advises a tiered s...

Endor Labs has launched Package Firewall, a new capability that blocks malicious, vulnerable, and non-compliant open-source packages before they reach developer machines or CI runners. The firewall sits between developer tools (including AI coding agents like Cursor and Claude Code), private registries, and public repositories, intercepting every install request. It uses a real-time malware feed updated within minutes of new package releases, with an AI-enhanced analysis pipeline that scans npm, PyPI, and Go ecosystems. The tool addresses the sharp rise in software supply chain attacks—including a 14x increase in OSV malware advisories over two years and 92% of npm account takeovers occurring in 2025. Recent incidents like backdoored versions of popular packages (lightning, axios, telnyx) highlight the need for immediate enforcement. Beyond malware, Package Firewall enforces organizational policies to block packages with known vulnerabilities or license violations and supports features...

As global cybersecurity spending approaches $240 billion and the average data breach cost hits a record $4.99 million, a growing affordability crisis is leaving small and midsize businesses (SMBs) dangerously exposed, threatening overall supply chain security. The market disproportionately serves large enterprises, with vendors developing expensive tools that don't account for SMBs' lack of dedicated security teams or 24/7 operations centers. AI adoption is adding cost pressure without guaranteed returns—research shows AI-generated patches fail or introduce new issues over 53% of the time. Security teams face tool sprawl (averaging 83 tools from 29 vendors), overlapping findings, and rising costs that outpace budget growth. Experts recommend shifting focus from raw findings to measurable risk reduction, applying cost governance to AI spending, and adopting frameworks similar to FinOps to ensure every dollar and engineering hour addresses the most critical exposures. The crisis ...

Zimperium's zLabs team has identified a major update to the ToxicPanda Android banking Trojan, which now features a significantly expanded command set of 167 remote commands and a dramatically broader global targeting scope. While the previous version targeted only 16 banking applications, ToxicPanda 2.0 now supports phishing overlays for 349 financial institutions, e-wallets, and cryptocurrency apps across 16 countries. The malware abuses Android Accessibility Services for remote control and has implemented new capabilities, including an automated process to enable Wireless Debugging (ADB) for privilege escalation to shell-level access, a PIN theft mechanism targeting over 140 applications, and the ability to steal lock screen credentials through overlay attacks. It also uses deceptive "system update" screens to hide malicious activity and delivers samples through Amazon AWS-hosted buckets. The malware communicates via encrypted WebSocket connections to its command-and-c...

This August 2026 report from CISA establishes a baseline of the vulnerability landscape before widespread AI-enabled discovery. It finds that most compromises stem not from sophisticated zero-day exploits, but from opportunistic criminals exploiting basic, preventable software weaknesses—with ransomware costing organizations an average of $3.7 million per incident. Analysis of CVE records shows a persistent recurrence of a small set of Common Weakness Enumerations (CWEs), with injection flaws, improper input validation, and memory safety issues dominating both the general CVE pool and CISA's Known Exploited Vulnerabilities (KEV) catalog. Critically, 41.5% of KEVs map to "stubborn weaknesses" documented for years, and some vulnerabilities considered "unforgivable" in 2007 persist today. The report introduces CISA's shift from relying on CVSS scores to a four-variable risk prioritization model (asset exposure, KEV status, exploit automation, and technical impa...

While 2026 is on pace for nearly 70,000 CVE disclosures - a record surge driven largely by GitHub's expanded advisory program and AI-assisted discovery - the actual exploitable risk has remained flat. Security expert Jerry Gamblin explains that most of the volume represents "old debt" surfacing through automated tooling rather than new dangers. The OWASP Top 10 vulnerabilities remain largely unchanged, and AI is simply finding the same common mistakes at scale. The breakdown of the NVD as a central enrichment authority is forcing a long-overdue shift of responsibility back to CNAs and vendors. Gamblin's key advice for practitioners is to treat CVE data as a paid product, using procurement leverage to demand better records from vendors. He emphasizes that genuinely exploited bugs still target familiar classes like VPN concentrators, and the real unsolved problem remains poor asset inventory - organizations cannot triage what they cannot see. The true AI-driven threat i...

After five years as an open-source company, Cal.com shifted to closed source in April 2026 to counter rising AI-driven security threats targeting its public code. However, this move didn't prevent new vulnerabilities from being introduced, especially as pull request volume grew from 40 to 100 daily with AI coding tools. The security team of two engineers struggled with noisy, conflicting results from multiple scanners, creating more reconciliation work than actual security. By adopting Gecko Security as a single source of truth in their pipeline, Cal.com gained context-aware scanning that provided working proofs of concept and prioritized real risk over raw severity. Gecko routes findings directly to code owners, stores triage comments to refine future fixes, and integrates security into the engineering workflow. This predictable, continuous process allows a 25-person team to achieve robust security coverage without slowing development, turning the challenge of AI-accelerated attac...

An analysis of 569 security vulnerabilities disclosed by n8n and GitLab reveals that most are not new issues but repeats of previously patched root causes. The study traced every disclosure back to its fixing commit, finding that 67% of n8n disclosures and 58% of GitLab CVEs re-fixed a known problem. These 569 findings condensed into just 105 underlying design decisions or "root causes," suggesting that fixing a class once at a shared point could prevent recurrence. The financial impact of these repeated fixes was significant, with over $1 million paid in bounties at GitLab alone, highlighting that recurring vulnerabilities represent a costly and often overlooked problem not captured by standard metrics.  https://www.gecko.security/lp/eliminating-vulnerability-classes