SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

OpenAI Uses AI to Block Unsafe Code Merges, but Blind Spot Risk Remains

0
·2 views

OpenAI has deployed a dedicated AI code-review model that automatically blocks pull requests flagged as security risks, with no human override allowed. Thibault Sottiaux, engineering lead for the Codex team, described the system on The Pragmatic Engineer podcast, claiming the model performs at a superhuman level in detecting logic errors and vulnerabilities. Beyond security, the same AI handles regression detection, dependency updates, and large-scale refactoring tasks that would otherwise take engineering teams months. However, critics point out a structural blind spot: if AI writes the code and AI reviews it, both systems may share identical failure modes. A vulnerability missed by the writing model could pass the review model undetected, creating a false sense of security that may be more dangerous than having no automated review at all.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Recursive CTEs Explained: The SQL Feature That Handles Unlimited Hierarchy Levels

Recursive Common Table Expressions (CTEs) solve a longstanding SQL limitation: traversing hierarchical data of unknown depth, such as org charts, category trees, or folder structures. A recursive CTE consists of two parts joined by UNION ALL — an anchor member that defines the starting row and a recursive member that repeatedly joins back to the CTE's own output until no new rows are returned. The technique works in PostgreSQL, SQL Server, and MySQL 8.0 or later, but is unsupported in older MySQL versions. Developers should use UNION ALL rather than UNION to avoid costly deduplication, and must guard against cyclic data — where relationships loop back on themselves — by tracking visited IDs or capping recursion depth. While powerful, recursive CTEs recompute the full traversal on every query, so alternative design patterns may be preferable for very large or frequently queried hierarchies.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

OpenProfile.md 0.2 lets users answer dating profile questions once, own the data

OpenProfile.md, an open Markdown standard served from a user's own domain, has released version 0.2 with a new Match section designed to eliminate repetitive data entry across dating platforms. The update allows users to fill in personal attributes — such as age, gender, orientation, height, and relationship preferences — a single time in a file they fully control. Any dating platform can then read the file directly, removing the need to re-enter the same information on Hinge, Bumble, or similar services. The specification includes a strict rule requiring a valid Born field confirming the user is 18 or older before a profile can appear in any matching context, and platforms are prohibited from inferring sensitive fields like orientation or ethnicity from external sources. Version 0.2 is published under CC BY 4.0 at logicsrc.com/docs/openprofile and is intended to be hosted at a well-known URL on the user's own domain.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer Builds Open-Source Tool to Protect BTS Fans from Ticket Scams

A developer and BTS fan has released 'sevensignal,' an open-source set of agent skills designed to help fans identify fraudulent ticket sellers and stay updated on official BTS news. The tool's ticket-safety checker flags any seller not explicitly named in a current BigHit or Weverse notice as 'unverified' rather than 'safe,' covering 41 known scam patterns including fake Weverse Shop clones, KakaoTalk fraud, and lookalike domains. A companion feature, official-news, summarizes BTS updates from public sources without requiring users to share their Weverse credentials. The project is MIT-licensed, entirely fan-made, and has no affiliation with HYBE, BigHit, or Weverse. It is publicly available on GitHub and open to community contributions, including new scam patterns and translations.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer builds AI 'skeptic' agent that found four major flaws in his own design doc

A developer building a SaaS product created a custom AI sub-agent explicitly programmed to refute rather than approve documents, and barred it from offering any positive closing remarks. When he turned the agent on his own finished design document, it surfaced four major issues he had overlooked. The flaws included a self-contradicting rule about blocking outbound communication, a legally unconfirmed claim written in conclusive language, a completion condition referencing an undefined ledger, and a circular dependency between project phases. The developer noted that standard AI review prompts tend to produce overly validating responses, while inverting the success condition — making the agent's goal to disprove rather than approve — forced more rigorous scrutiny. He concluded that authors are poorly placed to review their own work, and that an adversarially framed AI reviewer can surface blind spots that conventional review misses.

0 comments Read more at DEV Community