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
How to Prepare Your Website for Data Center Outages: Business Resilience Checklist
Following strikes on Yandex data centers, many businesses discovered that their sites and all backups were stored in a single location, leading to complete outages. The article provides a detailed checklist to mitigate such risks by separating critical components across different providers. Key recommendations include moving DNS to an independent provider, lowering TTL values, maintaining a live mirror site on separate hosting, and isolating corporate email. Additional steps cover local backups, monitoring, access management, and CDN usage. The guidance differentiates between virtual hosting, VPS or dedicated servers, and website builders like Tilda, while emphasizing compliance with Russian data localization laws such as 152-FZ.
How to Properly Compare CDNs: TTFB, p95 Latency, Caching, DDoS Protection and True Pricing
Selecting a CDN based solely on the number of points of presence is as unreliable as choosing a car by dashboard buttons. Networks with dozens of global PoPs may underperform for audiences concentrated in Novosibirsk or Yekaterinburg, while smaller regional providers often deliver superior peering with Russian operators. The article evaluates CDNVideo, EdgeCenter, Curator, NGENIX and Yandex Cloud CDN across verifiable features, network design, caching behavior, security controls, operations and commercial models. Performance data is separated because marketing claims cannot substitute for independent measurements from services such as CDNMON. Emphasis is placed on p95 metrics, cache hit ratios, origin shielding, raw logs and API limits rather than marketing scores. The analysis shows that no single universal ranking exists; optimal choice depends on traffic profile, audience geography, security requirements and operational workflows.
Innostage Lab Benchmarks PT NGFW 1.10.3 Performance in Eight Load Scenarios
Innostage researchers conducted extensive performance testing of PT NGFW version 1.10.3, the certified release at the time of trials. The evaluation covered TCP connection setup rates, concurrent session capacity, and throughput on a custom EMIX traffic profile with varying security modules enabled. Tests used Cisco TRex generators and a high-rule-count policy of 100,000 entries to simulate realistic enterprise conditions. Results showed the device sustaining 700,000 CPS and 6.4 million concurrent connections, matching or approaching vendor claims under HA-cluster configuration. Throughput remained above 27 Gbps even with IPS, application control, antivirus, NAT, URL filtering, and logging active. The report highlights differences in test methodology compared with vendor figures, including packet sizes and clustering setup.
Java Card Tutorial: Implementing Supplementary Security Domain to Limit ISD Privileges
A detailed technical guide demonstrates how to create a Supplementary Security Domain (SSD) on Java Card platforms to isolate applet management from the Issuer Security Domain (ISD). The approach prevents full card compromise if ISD keys are exposed by restricting SSD privileges to specific functions such as accepting extradition only from the ISD. Using the FunGP library and SCP02 protocol, the tutorial walks through installation, personalization with encrypted keys, and extradition of applets under the new domain. GlobalPlatform specifications including GP CS v2.3 and GP CIC v2.0 are followed for install parameters, privileges, and STORE DATA commands. After personalization the SSD reaches the PERSONALIZED lifecycle state, allowing secure delegation of applet control while the ISD retains overall card management rights. The method eliminates the need for custom operating systems and provides a standards-compliant way to enforce least-privilege access on smart cards.