Top LLMs Misidentify Poisonous Mushrooms in Every Ninth Case, Benchmark Shows
Developer Piotr Migdal tested whether current large language models can be trusted to identify mushrooms from photographs, focusing on the highest-risk step of foraging. The short answer is that relying on them could lead to serious poisoning incidents.
The benchmark used 1040 photographs of 55 mushroom species most common in Poland. All images were taken from the FungiTastic dataset, which itself is built on the Atlas of Danish Fungi. The full underlying collection contains more than 600,000 expert-labeled photographs, and a subset of specimens also carries DNA-based species confirmation.
Each model received one photograph at a time and was instructed to return the five most probable species names in Latin. No fine-tuning or external tools were provided. Gemini 3.8 Flash achieved the highest scores, placing the correct species first in 65 percent of cases and within the top five answers in 85 percent of cases. Gemini 3.6 Flash followed with 64 percent and 85 percent, while Gemini 3.7 Flash recorded 61 percent and 82 percent.
At the bottom of the accuracy table were Qwen 3.8 27B, MiniMax-M3, and Qwen 3.8 Flash. Accuracy figures dropped sharply when safety was considered. Both Gemini 3.8 Flash and Gemini 3.6 Flash classified poisonous mushrooms as edible species in approximately 11 percent of cases. The error rate rose to 24 percent for GPT-5.6 Sol, 29 percent for Claude Opus 5, and 36 percent for Qwen 3.8 27B.
Models were never asked directly whether a find was safe to eat. Instead, the researcher mapped the returned species names against a separate table of edibility. Some models refused to provide any classification at all for certain images.
Related articles
Developer Creates Personal Task Manager Allowing Claude Code and OpenCode Agents to Autonomously Execute Marked Tasks
A developer has built a personal task management system at shar8.app that integrates directly with AI coding agents including Claude Code, Codex CLI, and OpenCode through the Model Context Protocol. Tasks marked with an 'Agent' flag are automatically picked up by scheduled agent runs, executed locally on the user's machine, and results are written back into the task notes for human review. The system supports both browser-based OAuth authentication and persistent bearer tokens for headless cron or CLI usage, with strict separation between human-initiated and agent-initiated actions. Security measures include treating task text as untrusted data rather than instructions, disabling all tools by default, and requiring explicit allow-lists for any external actions such as web search. The implementation uses a custom MCP server exposing fifteen specific tools for listing, creating, editing, and delegating tasks, while a companion plugin for Claude Code handles scheduled execution via native system schedulers like launchd and systemd timers. The author emphasizes that agents never assign work to themselves and that final task closure always remains with the human user.
T-Bank Warns Customers of Payment and Transfer Outages
On November 3, T-Bank notified users that payments and transfers may fail due to ongoing technical issues. The bank placed an in-app message stating that the problems are being fixed. Clients confirmed the disruption through reports to Interfax, noting that the online banking interface displays the warning upon login. The outage affects only payments and transfers, while other banking services remain available. No details were provided regarding the root cause, the number of impacted users, or an estimated time for full recovery. Monitoring service DETECTOR404 also recorded the incident, aligning with user complaints and the bank's official notice.
Rostelecom Refactors InfoSec Careers: Matching Employee Skills to Evolving Business Priorities
Rostelecom's information security block developed a systematic approach to reassess employee expertise and redirect talent to new roles when business structures change rapidly. In one case, an architect hired for a specific team was reassigned after the unit was disbanded, initially moving to a similar position that did not fully align with his broader competencies. HR-business partner Olga Lapshina examined the specialist's full background, including prior work in audits, methodology development, regulatory documentation, university teaching, and scientific research, plus strong mentoring abilities with younger staff. These skills matched the needs of the methodology department, leading to a transfer supported by an individual development plan and a mentor via the internal Tochka Rosta platform. Within two months the employee contributed improvements that accelerated preliminary InfoSec audits by 30 percent and risk assessments by 20 percent while expanding threat scenario coverage by 10 percent. The initiative demonstrates Rostelecom's emphasis on horizontal career movement and retention of experienced professionals through structured career tracks.
How to Avoid Becoming the Hero of 'My Program Deleted 800 GB': Safe Recursive Deletion in Bash and Python
A recent Habr article examines dangerous patterns in recursive file deletion scripts that have caused massive data loss incidents, including an 800 GB wipe triggered by Telegram Desktop on Windows. The piece analyzes classic mistakes such as empty variables in rm commands and extra spaces that turned single paths into multiple destructive arguments, referencing the 2015 Steam Linux bug and the 2011 Bumblebee installer error. It provides hardened Bash templates using set -euo pipefail, ${VAR:?} checks, realpath validation, and --one-file-system flags to prevent deletion of forbidden directories or symlinked targets. The Python section introduces a safe_rmtree wrapper that enforces root boundaries, rejects symlinks, uses resolve(strict=True), and checks shutil.rmtree.avoids_symlink_attacks before proceeding. Additional recommendations include dry-run modes, marker files, temporary directory usage, and logging to make deletion operations auditable and reversible where possible.