MeridianMeridian

Top LLMs Misidentify Poisonous Mushrooms in Every Ninth Case, Benchmark Shows

Polish developer Piotr Migdal evaluated leading large language models on their ability to identify mushrooms from photographs, using a dataset of 1040 images covering 55 species common in Poland. The images came from the FungiTastic dataset derived from the Atlas of Danish Fungi, with expert labels and partial DNA confirmation. Models were asked to return the five most likely species names in Latin without additional training or tools. Gemini 3.8 Flash performed best with 65 percent top-1 accuracy and 85 percent top-5 accuracy, followed closely by other Gemini variants. However, safety-critical errors remained high: Gemini models labeled poisonous mushrooms as edible in roughly 11 percent of cases, while GPT-5.6 Sol reached 24 percent, Claude Opus 5 reached 29 percent, and Qwen 3.8 27B reached 36 percent. The study did not ask models directly whether a mushroom was edible; species identifications were later cross-checked against toxicity tables.

AntiMalwareTranslated from Russian

Read at source

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.

Claude Opus 5Gemini 3.8 FlashFungiTasticPiotr MigdalQwen 3.8 27B
ShareXFacebookLinkedInReddit
Meridian digestThe week's most consequential security stories, distilled into one briefing.See the digest

More in Other

Mail.ru Reduces Lifetime 1TB Cloud Storage Promo to Standard 8GB Quota for Select Users
Other

Mail.ru Reduces Lifetime 1TB Cloud Storage Promo to Standard 8GB Quota for Select Users

Mail.ru has begun notifying a small group of long-term Cloud Mail users that their complimentary 1TB storage, originally granted through a 2013-2014 app installation promotion, will be reduced to the standard 8GB free tier. Affected users receive personalized deadlines, with some facing storage limits starting 30 October and final deletion of excess files scheduled for 30 January 2027. The company confirmed the change impacts fewer than 0.5 percent of Cloud Mail users and stated that core free features remain available. Once the 8GB limit is exceeded, uploads, edits, and sharing are blocked while users have three months to download data or purchase additional space before files move to the trash for 14 days prior to permanent removal. This follows a similar reduction of lifetime 100GB storage performed by the service in 2024.

AntiMalwareOct 29, 2026Translated from Russian

Kimi K3 Open Model Surpasses Closed AI Leaders with Million-Token Context and 896 Experts
Other

Kimi K3 Open Model Surpasses Closed AI Leaders with Million-Token Context and 896 Experts

Kimi K3 is an open-weights model that outperforms closed systems from GPT and Anthropic on several independent benchmarks. Developers achieved this by scaling three dimensions simultaneously: one million tokens of context, deeper layers, and 896 Mixture-of-Experts specialists. The team replaced the standard KV-cache with a fixed-size state matrix updated via the delta-rule, solving numerical instability by clamping the forgetting coefficient alpha between e^-5 and 1. Sequential state updates were parallelized across chips, GPUs, and multi-GPU nodes using the open-source FlashKDA kernel built on CUTLASS. Inter-layer information flow was preserved through learned pseudo-queries that selectively retrieve relevant outputs from prior blocks rather than passing full activations. Load balancing across 896 routed experts uses per-expert quantile thresholds and a distributed histogram method to guarantee fair token allocation without full sorting.

HabrOct 29, 2026Translated from Russian

Innostage Completes Functional Testing of PT NGFW 1.10.3 Across 156 Criteria
Other

Innostage Completes Functional Testing of PT NGFW 1.10.3 Across 156 Criteria

Innostage has published the second part of its evaluation of PT NGFW version 1.10.3, focusing on functional capabilities rather than raw performance. The testing covered 156 individual checks grouped into categories such as basic technical features, firewalling, IDS/IPS, web filtering, malware protection, logging, routing, deployment, cryptographic controls and access management. Out of the total, 136 checks passed fully, seven passed with remarks, and 13 did not meet the predefined criteria. Key confirmed functions include high-availability clustering, centralized management via PT-MNGT, AD-based rule application, Site-to-Site VPN, TLS inspection, and Syslog integration with external security systems. Several advanced scenarios, including custom block pages, application risk classification, and centralized software distribution, were not confirmed under the test conditions. The report provides detailed protocol notes that help organizations map product capabilities against their specific infrastructure and regulatory requirements.

HabrOct 29, 2026Translated from Russian