March 2026 Summaries

On March 31, 2026, a supply chain attack compromised the npm package axios, a widely used HTTP client library in the JavaScript ecosystem, following a similar incident involving LiteLLM a week prior. The attack involved publishing a malicious version of axios (1.14.1) that included a new dependency, plain-crypto-js, which did not exist before the attack, enabling a remote access trojan with command-and-control capabilities. This trojan executed a multi-stage attack, including the deployment of platform-specific backdoors, allowing attackers persistent access to compromised systems. The incident was quickly detected by Comet, who responded by auditing repositories, scanning developer machines, and ensuring no production systems were affected, largely thanks to their use of committed lockfiles which prevented automatic resolution to the compromised version. The swift response and preventive measures highlighted the importance of security awareness, fast detection, and robust incident response processes, emphasizing the need for vigilant defenses in the open-source community against such high-value targets.

Opik has introduced a series of updates aimed at enhancing visibility, optimization, and customization in AI development workflows, featuring the new Opik-OpenClaw plugin that offers native observability for OpenClaw agents. This plugin captures comprehensive data on agent behavior, such as LLM calls and tool executions, providing insights into cost and performance metrics. Opik's Optimization Studio now simplifies the tracking and analysis of prompt optimization runs, offering detailed evaluations and configuration comparisons to improve performance with a complete audit trail. Additionally, custom dashboards have been improved, allowing for the creation of multi-project views with auto-save features, while expanded model and provider support, including Gemini 3.1 and Claude Sonnet 4.6, enhances tracing and evaluation capabilities. Usability improvements across the platform include error pattern grouping, full-text search for traces, and more accurate cost tracking, alongside community engagement opportunities through events and hackathons.

On March 24, 2026, the Python package litellm, essential for numerous AI agent frameworks, suffered a supply chain attack that compromised the security of millions of downloads by publishing malicious versions (1.82.7 and 1.82.8) to PyPI. The attacker accessed the maintainer's credentials, likely via a related GitHub Actions compromise, and introduced two sophisticated techniques to execute harmful code upon installation or Python environment startup. This attack targeted various credentials, including cloud, SSH, Kubernetes, and database configurations, encrypting and exfiltrating them to a malicious domain. Comet's response involved a comprehensive audit of their repositories, developer machine scans, and immediate credential rotations, ensuring that no production data was compromised. The attack highlighted the importance of several security practices, such as using lockfiles, pinning dependencies to exact versions, and auditing CI/CD secret scoping to prevent similar vulnerabilities. Following the incident, the compromised versions were removed from PyPI, and the current version was resolved to a safe state, with affected environments advised to recreate and rotate credentials.

Few-shot prompting enhances the performance of AI agents by providing them with 2-5 examples as a miniature dataset to follow, improving their ability to handle real-world inputs and reducing unpredictability. Unlike zero-shot or one-shot prompting, few-shot prompting involves giving multiple examples to define a pattern, which helps the model better understand tasks, ensure consistency, and produce structured outputs. This technique is particularly valuable in agentic systems, where various smaller prompts power different steps in a workflow, such as interpreting messy user requests or mapping text to structured parameters. By using realistic examples, few-shot prompting addresses issues like tool-calling precision, structured output enforcement, and edge case handling, leading to more reliable agent behavior. The method avoids the need for extensive fine-tuning and enables faster iteration, with lower costs associated with errors. Implementing few-shot prompting effectively involves selecting diverse, production-realistic examples while managing the trade-off between token cost and performance improvement. The process can be optimized further using tools like Opik's Few-Shot Bayesian Optimizer, which helps find the best example combinations to enhance task performance while considering quality and cost.

OpenClaw, a rapidly growing open-source project, enables users to connect large language models (LLMs) to a variety of tools and applications, offering powerful capabilities for autonomous personal AI agents. However, it has faced criticism for lacking visibility into its operations, leading to concerns about token usage, task failures, and overall reliability. To address this, the opik-openclaw plugin has been introduced, providing native observability by capturing detailed traces of each interaction, including LLM calls, tool executions, and cost breakdowns. Developed by Comet, the plugin integrates seamlessly with OpenClaw, allowing users to monitor and evaluate AI agent performance with enhanced transparency and control. This solution promises to offer enterprise-grade reliability by facilitating full visibility into the agent's reasoning processes, without relying on network proxies or manual configuration.

At Comet, agent observability is often sidelined due to the fast-paced nature of feature development, prompting the creation of Opik to enhance observability in AI systems. Traditional testing methods fall short with nondeterministic LLM outputs, so Comet's engineering team developed the Opik Claude Code Plugin to automate observability and best practices, packaging their internal workflow into a tool that facilitates reliable AI agent development. The plugin's auto-instrumentation feature simplifies tracing for Python and JavaScript agents, aligning traces with logical execution steps rather than merely the code structure. This is particularly crucial for agentic systems, which operate with increasing complexity and autonomy, necessitating robust observability for performance and behavioral insights. The Claude Code tracing capability provides transparency into AI operations, capturing workflow details such as tool usage and costs, thereby transforming Claude Code from a simple assistant into a comprehensive agent engineering framework. This approach ensures that observability and best practices are seamlessly integrated into development processes, reducing technical debt and enhancing system reliability and efficiency.