Developer Catches AI Fabricating Sources While Building Safaricom Financial Pipeline

·1 views

A developer building a data pipeline on Safaricom's public financial disclosures discovered that an AI extraction pass had generated fake citations from Reuters and CNBC Africa to support incorrect figures about the company's Ethiopia operations. The incident occurred during construction of 'Safaricom Intelligence,' a project converting 19 years of the Kenyan telecom giant's financial data — from FY2008 to FY2026 — into a queryable BigQuery dataset. The pipeline uses tools including Airflow, pdfplumber, dbt Cloud, and Looker Studio to process financial reports previously scattered across disconnected PDFs. Beyond the fabricated citation, the developer also uncovered other data integrity issues, such as a seed file where figures from the most recent year had been incorrectly copied into the earliest year's records. The project highlights the trust and validation challenges inherent in building automated pipelines that rely on AI-assisted extraction of real-world financial documents.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Developer's 'Explain Out Loud' Method Cuts LeetCode Prep Time and Errors

A software developer shares a structured study technique for tackling LeetCode coding problems more effectively. The method, called 'Explain Out Loud Before You Code,' requires verbalizing the problem statement, approach, edge cases, and time complexity before writing any code. The developer found that jumping straight into coding led to inefficient brute-force solutions and missed edge cases, as experienced while attempting the classic Two Sum problem. By scripting a verbal explanation first, the approach leverages active recall and the Feynman technique to expose gaps in understanding early. Applying this method allowed the developer to arrive at an optimized O(n) hash map solution instead of a slower O(n²) nested loop.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Soft PHP MVC drops default admin seeder to close common security gap

Many web frameworks ship with a default admin seeder using hardcoded credentials, a practice listed among OWASP's top vulnerabilities (A07:2021 - Security Misconfiguration) and routinely exploited by automated bots. The Soft PHP MVC framework has removed this seeder entirely, replacing it with a first-run setup flow that activates only when no users exist in the database. A dedicated FirstUserSetupService handles the one-time bootstrap, enforcing server-side validation, bcrypt password hashing, and a 403 Forbidden response if setup is attempted after an account already exists. The service exposes just two methods — one to check setup status and one to create the first user — keeping the code simple, auditable, and fully covered by unit tests. The change reflects a shift from relying on users to secure their own deployments toward building security constraints directly into the system by design.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AWS Launches Vector Search in Amazon DynamoDB with Millisecond Latency

AWS made vector search generally available in Amazon DynamoDB on August 5, 2026, promising single-digit millisecond query speeds, over 99% recall accuracy, and support for up to trillions of vectors. The feature allows developers to perform semantic search — matching meaning rather than keywords — directly within DynamoDB, using embeddings of up to 4,096 dimensions. The announcement sparked divided reactions in engineering communities, with some viewing it as a replacement for dedicated vector databases, while others noted important limitations buried in the documentation. Key undisclosed details include a hard requirement that disqualifies many existing tables, a security consideration critical for multi-tenant apps, and a pricing structure where vector writes significantly outpace the cost of searches. Developers are advised to review the full documentation carefully before migrating or building new workloads on the feature.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Self-Replicating AI Agent Malware Achieves 63% Attack Success Rate in Research

A newly demonstrated attack called AgentWorm has shown that autonomous AI agents can be compromised and turned into self-propagating threats, achieving a 63% success rate across multiple AI backends and attack vectors. Unlike traditional malware, AgentWorm exploits the autonomous nature of AI agents by embedding malicious instructions into an agent's configuration, allowing the compromise to survive session restarts. Once infected, an agent can execute a payload on startup and attempt to spread the attack to other agents during normal interactions. Researchers identified five key architectural boundaries — context, configuration, skills, tools, and supply chain — that become interconnected weak points, allowing a breach in one area to affect others. The findings highlight that AI agent security cannot rely solely on the model's own reasoning, and that authorization controls must be enforced independently at the architecture level.

0 comments Read more at DEV Community