September 2026 Summaries

Claude Opus 5.5 was evaluated on all 186 SusVibes tasks, which reconstruct real Python-project features associated with public CVEs and scores both functional correctness and resistance to the original vulnerability; it produced working code for 93.5% of tasks and correct, secure code for 54.8%, which would lead the benchmark’s public leaderboard at the time of testing. However, analysis found that 53% of its solutions, including 56% of secure solutions, closely resembled the original project code, suggesting that training-data memorization substantially contributed to its results despite the model having no network or repository access during evaluation. Secure outcomes were notably stronger for memorized code than independently generated code, particularly on large fixes, while recalled pre-fix implementations sometimes reproduced vulnerabilities. Performance also differed sharply by weakness type: the model did well on cross-site scripting and incorrect authorization tasks but failed to securely solve many tasks involving cryptography, concurrency, authentication, injections beyond SQL, and browser or session protections. The findings argue that public-CVE benchmarks may increasingly measure familiarity with published fixes rather than general secure-coding ability, and recommend treating passing tests separately from security assurance, closely reviewing complex changes and high-risk areas, and examining solution similarity to reference implementations when interpreting benchmark rankings.

Application security (AppSec) focuses on protecting the software an organization builds, including its code, infrastructure, libraries, components, and delivery pipelines, primarily through collaboration with developers rather than reliance on tools alone. Its broad responsibilities include secure design, developer training, vulnerability discovery and remediation, supply-chain security, security metrics, and integration of controls throughout the software development lifecycle, while activities such as endpoint patching, general employee awareness training, threat monitoring, and vulnerability research are usually handled elsewhere. Effective AppSec programs rest on people, processes, and tools, with developer relationships and workflow integration considered more important than scanners or platforms. Core technologies include static application security testing (SAST), which analyzes code before it runs; dynamic application security testing (DAST), which evaluates running applications; and software composition analysis (SCA), which inventories dependencies and identifies known vulnerabilities, malicious packages, and exposed secrets. Faster release cycles, widespread open-source dependencies, custom CI/CD pipelines, and AI-generated code have increased both the amount of software and the difficulty of securing it. AI can improve triage, detection, and embedded developer security assistance, but it also expands attackers’ capabilities and creates greater volumes of potentially vulnerable code, reinforcing the need for prioritization, secure design, and human judgment around business logic and risk.

Semgrep Supply Chain has introduced Malware Detection and Response Automation to help organizations identify, respond to, and eventually prevent malicious dependency incidents across their software repositories. Its zero-day detection capability uses AI-assisted, expert-verified rules to create incident advisories within minutes, automatically rescan customer codebases, identify affected projects and blast radius, and notify opted-in customers through Slack, webhooks, and dashboard alerts. Organization-wide policies can then automatically initiate actions such as creating Jira tickets, sending SecOps notifications, invoking webhooks, and rescanning repositories after remediation. The announcement responds to growing package ecosystem threats, with an estimated 1.8 million malicious packages reported in the second quarter of 2026. Semgrep is also developing a private-beta Malware Firewall integrated with Guardian that intercepts package requests between package managers and public registries, blocking known malicious dependencies before they reach developer machines without requiring changes to developer workflows.

A reported compromise of MemTensor’s MemOS AI memory framework illustrates a supply chain attack that embeds credential-stealing malware in normal runtime paths rather than relying on installation scripts or typosquatting. Malicious releases published on September 23, 2026 affected the npm package @memtensor/memos-cloud-openclaw-plugin versions 0.1.21, 0.1.23, and 0.1.25 and the PyPI package MemoryOS version 2.0.34, with the Python payload activating during package import and the npm payload activating when OpenClaw starts or recalls memory. The malware, known as sckit, searches hosts for credentials, environment variables, registry tokens, cloud and source-control keys, Hugging Face tokens, and potentially prompt content, then sends collected information to command-and-control infrastructure. The incident highlights the heightened risk to AI development tooling, where systems often hold numerous credentials and may receive less supply-chain scrutiny than traditional software dependencies. Organizations that used the affected versions are advised to treat impacted machines and CI runners as compromised, remove or downgrade the packages, rotate accessible secrets, investigate unexpected package releases, remove malware artifacts, block associated infrastructure, and adopt dependency inventory, trusted publishing, and provenance verification practices.

Semgrep’s summer interns worked on production engineering challenges spanning infrastructure scaling, product integrations, program analysis, and internal performance tooling. Tanisha Kholiya improved Celery queue autoscaling through load testing and iterative tuning, reducing average backlogs by about 99% while lowering pod costs roughly 30%. Isabel Berny developed automatic Jira status synchronization and finding-level subtasks, while also expanding a small repository-tag task into a broader product and design effort around preventing accidental policy changes. Solomon Graf replaced a heuristic variable-origin check with reaching-definitions analysis, balancing programming-languages theory with a narrower implementation suited to production needs. Xavier Lien added memory profiling to Semgrep’s OCaml profiler, which helped identify an OCaml compiler bug and reduce a customer’s scan memory usage by approximately tenfold. Across the internships, the company emphasized ownership, mentorship, cross-functional collaboration, exposure to customer and incident work, and social activities, while announcing applications for its 2027 software engineering internship.

Semgrep CTO Drew Dennison’s Black Hat 2026 talk argued that traditional “shift left” security practices remain necessary but must be complemented by “shifting right” to continuously analyze deployed and legacy software as AI expands both code production and attacker capabilities. Semgrep researchers, using early access to Anthropic’s Mythos model alongside static analysis, dynamic testing, and constrained agent skills, reported finding 100 runtime-verified vulnerabilities in major open-source projects, emphasizing that specialized analysis tools and testing harnesses can be as important as model capability. The company’s Mandoline project is designed to give AI agents structured program-analysis access across 47 languages, reducing repeated text searches and file reads by supplying code graphs, dataflow analysis, taint tracking, reachability information, and audit coverage. Its Context Engine aims to make recurring security analysis affordable by caching prior conclusions and retaining organizational knowledge, such as accepted risks and past triage decisions, so agents do not repeatedly flag known nonissues. The central argument is that defenders can offset attackers’ growing access to capable, cheaper AI models by combining continuous analysis with proprietary codebase, deployment, threat-model, and institutional context, while also applying security controls earlier in AI-assisted coding workflows.