July 2026 Summaries

Every day, new Common Vulnerabilities and Exposures (CVEs) and security advisories emerge, and the Semgrep supply-chain team examines them, identifying patterns in cryptographic advisories. These patterns often reveal that vulnerabilities arise not from the cryptographic math itself but from skipped validation checks in the code, such as improper signature verification and algorithm confusion. This insight led to an extensive analysis of the GitHub Security Advisory (GHSA) dataset, confirming that many cryptographic failures involve code errors, not mathematical ones. A real-world example is CVE-2021-43572, where a FinTech signing library failed to perform a simple range check on ECDSA signature values, allowing attackers to bypass signature verification. This recurring issue highlights the need for focusing on code validation rather than cryptography itself, as many security bugs arise from overlooked input checks. Presentations at BSidesLV and DEF CON 34 emphasize that even without a deep cryptography background, understanding what an attacker can manipulate and applying range checks can prevent these vulnerabilities.

Semgrep has launched nine pre-built Agentic Workflows, currently in public beta, to enhance application security by quickly detecting complex flaws such as authentication, injection, and business logic errors. These workflows utilize the Semgrep Pro Engine, AI models like Claude Opus, and Semgrep Mandoline, a code slicer, to perform deterministic analysis and reasoning about exploitability, effectively automating the identification of vulnerabilities at a speed and scale beyond traditional methods. Designed to counter the advanced capabilities that attackers now possess, these workflows enable continuous security monitoring of codebases, offering AppSec teams a robust tool to match the increasing pace of AI-generated code reviews. Initial testing with design partners showed high accuracy in detecting true positives, aligning closely with internal review results, and the platform supports both out-of-the-box use and further customization with a Python SDK and CLI. As Semgrep continues to develop this technology, future enhancements will focus on triage, remediation, and proactive security measures to further strengthen AppSec processes.

Open weight and open source models have been gaining attention for their ability to challenge benchmark records and their potential as security tools, though they face scrutiny in the U.S. for security risks. These models offer control over data residency and cost-effectiveness compared to frontier models like Anthropic's Claude and OpenAI's GPT, but their lack of guardrails might lead to regulatory restrictions. Kimi K3, a recent open weight model, did not outperform frontier models in benchmarks for detecting vulnerabilities such as Insecure Direct Object References (IDORs), especially in large, interconnected codebases, due to its lower precision and recall rates. While Kimi K3 may suit smaller projects, it requires more triage effort and is not a seamless substitute for models like GLM in larger environments. Security teams are advised to test models against representative repositories and consider their triage capabilities when choosing models, as the choice of model can significantly impact efficiency and security outcomes.

Hacker Summer Camp, encompassing Black Hat, BSidesLV, and DEF CON, is set to feature a significant presence from Semgrep, with numerous talks and events highlighting topics like AI-assisted development and cryptographic failures. Semgrep's participation includes sessions on AI slop detection, hackable eReaders, and real-world cryptographic failures, presented by key figures such as Diptendu Kar and Katie Paxton-Fear. The company will also host a booth at the Black Hat Business Hall, offering discussions on AI integration in security, and panels with industry experts like Leif Dreizler and Cathy Polinsky. Additionally, the event promises engaging interactions beyond scheduled talks, encouraging attendees to engage in conversations, explore live demos, and participate in gaming nights.

A new wave of open-source AI security projects is emerging, focusing on utilizing large language models (LLMs) to identify vulnerabilities in codebases. These tools fall into three main categories: LLM-led exploit generation, LLM-skill-boosting, and hybrids combining static application security testing (SAST) with LLMs. LLM-led exploit generation tools aim to identify vulnerabilities by driving code to a crash state, akin to fuzz testing, although practical application is limited due to model guardrails and narrow vulnerability focus. LLM-skill-boosting tools enhance LLMs with reasoning capabilities to mimic human vulnerability researchers, while SAST+LLM hybrids employ deterministic analysis tools alongside LLMs to refine the search for vulnerabilities. Despite overlapping functionalities and uncertain maintenance futures, each tool offers specific advantages for distinct use cases, such as local offline scanning, high-confidence findings for C/C++ maintainers, or streamlined app security programs. The field is rapidly evolving, and no clear market leader has yet emerged, with many companies likely to develop bespoke solutions suited to their needs.

The blog post explores the effectiveness of large language models (LLMs) in detecting Insecure Direct Object Reference (IDOR) vulnerabilities within codebases through a benchmarking study. It evaluates the models using traditional metrics like precision, recall, and F1 score, but introduces additional measures such as groundedness, counterfactual reasoning, metamorphic testing, selectivity, and stability to assess the models' reasoning capabilities. The findings reveal that while none of the models are exploiting the benchmark through pattern matching, they all share a significant weakness in recall, struggling to identify a majority of the vulnerabilities. The study highlights that the models are effective at identifying straightforward vulnerabilities but fail to detect more complex issues, pointing to a common challenge across the field. The post emphasizes that the real frontier in this domain is improving recall to capture the vulnerabilities that currently go unnoticed by all models, including the author's own Semgrep Multimodal agent.

In a sophisticated cyberattack on July 14, 2026, an attacker gained access to the asyncapi/generator repository, pushing a single malicious commit that triggered the project's GitHub Actions workflow, leading to the release of tampered packages on npm with valid provenance attestations. The malicious code, hidden through obfuscation and padding, activates upon module use, not installation, indicating that npm's security measures are influencing attacker strategies. The attack utilized a Remote Access Trojan (RAT) named "Miasma v3," capable of credential harvesting, lateral movement, and AI coding-assistant poisoning, while employing multiple command-and-control channels for resilience. Despite the attack's complexity, the authors deliberately disabled npm propagation to avoid detection, underscoring the increased professionalism and evolving tactics in cyber threats. This incident highlights the importance of continuous vigilance and adaptation in security practices, as enhanced security measures like script blocking, while effective, are not foolproof solutions.

OpenAI's release of the GPT-5.6 system card marks the introduction of three new models—Sol, Terra, and Luna—that meet high cybersecurity capability thresholds, streamlining the automation of end-to-end attacks and vulnerability discovery. While these models show a regression in precision, they offer a significant improvement in recall on true positives, with Luna standing out for its cost-effectiveness. The performance of these models is further enhanced when integrated with comprehensive security workflows involving tools like Semgrep Guardian, rather than relying on raw code analysis alone. While benchmarks reveal improvements over previous versions like GPT-5.5, the results also highlight the importance of the surrounding security infrastructure in optimizing model use. Organizations must carefully consider the trade-offs between using open-source models and commercial offerings, as each has implications for cost and risk management, especially concerning the gap between code generation and review. With the cybersecurity landscape continuously evolving, leveraging models like GPT-5.6, along with other tools, is crucial for maintaining a balanced defensive and offensive posture in AI-driven environments.

Modern AI models, including open-source and proprietary frontier models, present significant challenges in terms of transparency and trust, as their inner workings remain largely inscrutable compared to traditional software. This lack of mechanistic interpretability makes it difficult to predict model behavior or detect potential backdoors—malicious manipulations that can subtly influence outputs—introduced during training or fine-tuning. While no public evidence currently suggests widespread deliberate poisoning of open-source models, the risk of such compromises remains a concern, as they can lead to biased recommendations or insecure outputs that are hard to trace back to the model itself. The industry's reliance on benchmarks and marketing claims for assessing model reliability is insufficient; instead, a robust ecosystem involving trusted third parties for independent evaluation, provenance tracking, and transparency at every development stage is necessary to build trust in AI technologies. This situation mirrors Ken Thompson's insights on the importance of understanding the full lineage and development process of systems to ensure their security and reliability.

Hacker Summer Camp is an annual event in Las Vegas that draws hackers worldwide for three consecutive conferences: BlackHat, BSides, and DEF CON. BlackHat is a corporate-focused event featuring a Business Hall with cybersecurity vendors, Arsenal for open-source tools, and technical talks. BSides serves as a community-driven counterpart to BlackHat, offering a more approachable atmosphere with various "grounds" for networking and knowledge-sharing, including off-the-record Skytalks. DEF CON is a chaotic, community-driven hacker conference known for its villages, where attendees can engage in interactive sessions and workshops. The author advises attendees to balance their time across these events, prioritize networking, and be open to new experiences while navigating Las Vegas's unique environment. Emphasizing the importance of taking breaks, staying hydrated, and engaging in informal networking, the author highlights the significance of embracing the diverse and sometimes unconventional aspects of the event to maximize the experience.

In 2024, the FTC initiated fines for companies falsely claiming to use AI, targeting deceptive practices like "AI-powered" tools that had little to no real AI integration. This crackdown highlights the trend of "AI-washing," where products are superficially labeled as AI-enhanced without substantial changes to their underlying technology. The text discusses the importance of distinguishing between genuinely AI-native tools, which have AI integrated into their core functions such as detection and analysis, and those that merely use AI for superficial features like user interfaces. It emphasizes the need for rigorous evaluation of AI claims, particularly in application security, where AI-native solutions are essential for addressing the unique challenges posed by AI-generated code. The text advises asking specific questions to determine if AI is integral to a product's architecture or merely a decorative feature, encouraging buyers to ensure that AI truly enhances the tool's capabilities rather than just serving as a marketing gimmick.