#MLSecOps
"TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering", Feb. 2026.

]-> Toolkit to benchmark the tamper-resistance of LLMs

// As increasingly capable open-weight LLMs are deployed, improving their tamper resistance against unsafe modifications, whether accidental or intentional, becomes critical to minimize risks. Varied data sets, metrics, and tampering configurations make it difficult to compare safety, utility, and robustness across different models and defenses. We introduce TamperBench, the first unified framework to systematically evaluate the tamper resistance of LLMs

ИСП РАН может стать ведущим участником единой системы сертификационного аудита для русскоязычных ИИ-моделей

Выбор Института системного программирования им. В.П. Иванникова РАН объясняется технической экспертизой организации. Платформа для сертификации может быть создана в рамках Исследовательского центра доверенного искусственного интеллекта ИСП РАН.

Новая система будет оценивать соответствие нейросетей техническим и нормативно-этическим стандартам. Модели проверят на уязвимости, устойчивость к атакам, на отсутствие манипулятивных алгоритмов.

Аудит затронет все этапы жизненного цикла продукта, в том числе разработку и эксплуатацию. В первую очередь планируется проверить решения для государственного сектора и объектов критической информационной инфраструктуры. Также будет изучаться работа ИИ в высокорисковых сферах, таких как медицина и финансы.

🔗Источник: https://iz.ru/2038342/2026-02-06/isp-ran-mozhet-zaniatsia-sertifikatciei-tekhnologii-iskusstvennogo-intellekta-v-rf

⚡️ X-Coder - новый мощный релиз для код-LLM от IIGroup.

Команда выложила полноценный стек для обучения моделей программированию: синтетические датасеты, RL-выравнивание и готовые модели с сильным логическим мышлением.

Что внутри:

— X-Coder-SFT-376k - большой синтетический датасет (376K примеров), сгенерированный с помощью GPT-o3-mini и DeepSeek-R1
— X-Coder-RL-40k - набор для RL-дообучения, сфокусированный на логике, проверке решений и качестве рассуждений
— X-Coder-8B - модели на базе Qwen3 с упором на reasoning и задачи программирования

Главная идея —-полностью синтетическая эволюция данных: масштабирование качества без ручной разметки. Такой подход уже становится стандартом для обучения сильных coding-моделей.

Полезно, если вы:

• обучаете собственные code-LLM
• исследуете synthetic data + RLHF/RLAIF
• работаете с Qwen-экосистемой или агентами для программирования

Модели и датасеты:
https://modelscope.cn/datasets/IIGroup/X-Coder-SFT-376k

https://modelscope.cn/datasets/IIGroup/X-Coder-RL-40k

https://modelscope.cn/models/IIGroup/X-Coder-SFT-Qwen3-8B

#LLM #CodeAI #Qwen3 #DeepSeek #AI #Coding

#Research
#Threat_Research
#Malware_analysis
"SoK: Exposing the Generation and Detection Gaps in LLM-Generated Phishing Through Examination of Generation Methods, Content Characteristics, and Countermeasures", Nov. 2025.

// This study presents a nine-stage taxonomy detailing how attackers exploit LLMs to generate phishing content while bypassing safety guardrails. Findings indicate that these AI-generated threats effectively evade detection by leveraging human cognitive manipulation, highlighting a critical asymmetry where offensive tactics evolve dynamically compared to reactive, static defenses

OWASP Vendor Evaluation Criteria for AI Red Teaming Providers & Tooling v1.0

https://genai.owasp.org/resource/owasp-vendor-evaluation-criteria-for-ai-red-teaming-providers-tooling-v1-0/

#AIOps
#Research
"When Benign Inputs Lead to Severe Harms:
Eliciting Unsafe Unintended Behaviors of Computer-Use Agents", Feb. 2026.

]-> AutoElicit Datasets
]-> Code

// The first conceptual and methodological framework for studying unintended behaviors of CUAs: unsafe actions that emerge inadvertently from benign inputs during typical user interactions

#Analytics
#Research
#MLSecOps
"Trojans in Artificial Intelligence. Final Report",
IARPA, Jan. 2026.

// This report synthesizes the TrojAI program's key findings, including methodologies for detection through weight analysis and trigger inversion, as well as approaches for mitigating Trojan risks in deployed models. Comprehensive test and evaluation results highlight detector performance, sensitivity, and the prevalence of "natural" Trojans. The report concludes with lessons learned and recommendations for advancing AI security research

MITRE ATLAS OpenClaw Investigation Discovers New and Likeliest Techniques - https://www.mitre.org/sites/default/files/2026-02/PR-26-00176-1-MITRE-ATLAS-OpenClaw-Investigation.pdf

MITRE ATLAS™ analyzed OpenClaw incidents that showcase how AI-first ecosystems introduce new exploit execution paths. OpenClaw is unique because it can independently make decisions, take actions, and complete tasks without continuous human oversight.

AI Security Guide and Risk Assessment Tool - https://www.rand.org/pubs/tools/TLA4174-1/ai-security/guide.html by RAND

This guide is a practical, risk-based resource for developers, security experts, and policy professionals navigating the AI security landscape.⁠1 The guide addresses security of AI systems broadly, including machine learning (ML) models and other AI-enabled architectures. Certain sections, such as the threat landscape and model weight protection sections, focus more specifically on statistical, ML-based models. Building on industry best practices and expert insights, the guide helps you understand and manage the security risks associated with AI systems across their lifecycle—from design and development to deployment and operation.

secureclaw - Automated security hardening for OpenClaw AI agents - https://github.com/adversa-ai/secureclaw

51 audit checks. 12 behavioral rules. 9 scripts. 4 pattern databases. Full OWASP ASI Top 10 coverage.

SecureClaw audits your OpenClaw installation for misconfigurations and known vulnerabilities, applies automated hardening fixes, and gives your agent behavioral security rules that protect against prompt injection, credential theft, supply chain attacks, and privacy leaks.

What Problem Does SecureClaw Solve?

AI agents with access to your files, credentials, email, and the internet are a fundamentally different security surface than traditional software. An agent that can read your .env file and send HTTP requests can exfiltrate your API keys in a single tool call. An agent that trusts instructions embedded in a web page or email can be hijacked to act against your interests.

SecureClaw addresses this by operating on three layers:

Layer 1 - Audit. 51 automated checks across 8 categories scan your OpenClaw installation for known misconfigurations: exposed gateway ports, weak file permissions, missing authentication, plaintext credentials outside .env, disabled sandboxing, and more.

Layer 2 -Hardening. Automated fixes for the most critical findings: binding the gateway to localhost, locking down file permissions, adding privacy and injection-awareness directives to your agent's core identity file, and creating cryptographic baselines for tamper detection.

Layer 3 - Behavioral rules. 12 rules loaded into your agent's context that govern how it handles external content, credentials, destructive commands, privacy, and inter-agent communication. These rules cost approximately 1,150 tokens of context window and provide defense against prompt injection, data exfiltration, and social engineering -- attacks that cannot be prevented by infrastructure configuration alone.