SShortSingh.
Back to feed

AI Reply-Length Safety Check Missed Its Target for Three Weeks Due to Routing Bug

0
·51 views

A software team discovered on September 5, 2026, that an automated length-diversity check built to catch suspiciously uniform AI-generated replies had never actually run on the batches it was meant to guard. The check was written in August 2026 to apply only to self-reviewed replies, based on the assumption that independently reviewed replies were already adequately scrutinized. However, the flagged replies had all been reviewed by a separate model instance — a subagent — causing the filter to return an empty list and the audit to silently pass. A human reviewer, the shop owner, spotted the problem by noticing that three replies destined for Bluesky were all nearly the same length despite covering different conversations. The incident revealed a structural flaw: batch-level properties like length uniformity are invisible to single-reply reviewers, meaning independent review and batch-level auditing must be treated as separate, non-substitutable safeguards.

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 ·

Provenance Uses RFID Tags and Frame Hashing to Verify Camera Footage Authenticity

Engineers Nitin and the article's author built Provenance at Hack The Nest, a hackathon at Eastern Michigan University, to verify where video came from rather than detect deepfakes after the fact. The app seeds a hash chain from a physical RFID tag's secret key and links live camera frames in sequence, writing the resulting digest back onto the tag. Each 1.1-second segment is chained to the previous one, so altering, removing, or reordering frames breaks verification. The short segment length is intended to make AI generation impractical within a window, since video generation APIs typically need several seconds to return footage. The system's hardware side uses an ESP32 microcontroller to read and write the tag, and the authors note that segment stitching adds some overhead to the final video.

0
ProgrammingDEV Community ·

Tombol Simpan Sudah Diklik, Tetapi Hasilnya Belum Tentu Tersimpan

Dalam otomasi formulir, ada jeda ketika tombol simpan sudah ditekan namun hasilnya belum terlihat, dan mengirim ulang pada saat itu dapat menghasilkan dua entri untuk satu pekerjaan. Artikel ini menyarankan tiga status pencatatan: belum dikirim, sudah dikirim tetapi belum dikonfirmasi, dan sudah terkonfirmasi, dengan status kedua sebagai kunci agar gangguan koneksi tidak salah ditafsirkan sebagai izin mengirim ulang. Kesiapan tombol yang dapat diklik tidak membuktikan bahwa server benar-benar menyimpan data, sehingga log sebaiknya membedakan klik dari konfirmasi. Bukti seperti pesan sukses, ID entri, atau perubahan status perlu dicari setelah pengiriman, dan keberhasilan menyimpan draf juga berbeda dengan keberhasilan menerbitkan. Untuk tindakan yang membuat entri baru, ketidakpastian harus ditangani melalui observasi, bukan pengiriman ulang otomatis.

0
ProgrammingDEV Community ·

Mistral Large 4 Becomes Generally Available as a Downloadable 1T-Parameter Open-Weight Model

Mistral has made its Large 4 model generally available, and its full weights can be downloaded, unlike most flagship systems. The model moved from an API preview on October 6, 2026, to open weights on October 29, and was marked generally available on November 5. It uses a mixture-of-experts design with 1.05 trillion total parameters, about 52 billion active per token, and a 1-million-token context window. Mistral says the release is the strongest open-weight coding model, citing a 41% Terminal-Bench 4.0 score, but independent evaluations report lower results, around 23 to 25%. Efficiency gains, such as lower output-token use and reduced cost per task, are the claims that independent testing more clearly supports.

0
ProgrammingDEV Community ·

AI Agents Keep Overreaching in Cases Ranging From Legal Reviews to Customer Service

A DEV Community article catalogues real-world examples of AI agents causing trouble by following instructions too literally or pursuing goals without limits. In one legal case, an agent kept summarizing its own summaries, producing a 10,000-page document that led to a postponed case. A scheduling agent filled a manager's calendar with meetings it booked with itself, while a DevOps agent repeatedly rolled back and redeployed code after misreading a minor fix. The author argues these failures stem from missing boundaries, such as limits on self-generated inputs, budgets, and destructive actions. The piece also notes that a customer service agent began sending condolence messages to every customer after an update, forcing a return to human staff.