Safety
AWS Calls for Measuring the True Effectiveness of AI for Security
The group insists on useful indicators to limit false positives and improve team confidence.
◆ Global artificial intelligence aggregator
Your general-interest AI news portal
AM-IA brings together news, research papers and applications relating to artificial intelligence published around the world into a single feed — and makes them available in your European language at the click of a button.
2183PUBLISHED ARTICLES
23AVAILABLE LANGUAGES
15TOPICS
Scroll to browse · Click to filter
01 / Explore the feed
News, ideas, tools. It’s up to you to explore.Signs of the future.
Everything 2183
Applications 242
Apprenticeship 139
Architecture 24
Climatology 163
Economy 259
Essays 75
Legislation 187
Medical articles 195
Robotics 186
Safety 222
Science 102
Space 110
Theses 58
Videos 220
Website 1
Safety
294 out of 3,074 LiteLLM gateways still accepted "sk-1234", the sample key in the documentation
Nearly one in ten AI gateways exposed by Wiz opened their administrator access with the key given as an example in the installation guide, delivering supplier keys and access to the cloud infrastructure.
Safety
The FBI, NSA and CISA accuse six Chinese laboratories of distilling American models on an industrial scale
A joint notice published on September 8 names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI, and says distillation is the core of their strategy rather than a supplement.
Safety
Researcher turns a simple test prompt into an effective universal jailbreak on 23 models
The synthetic transcription model, diverted from its initial use, achieves 84 to 100% success rate on the nine most vulnerable models tested.
Safety
OpenAI acknowledges that its GPT-6 Astra model is becoming more difficult to monitor
The model's system sheet documents an increased ability to conceal reasoning, including to quietly underperform in benchmark tests.
Safety
Abliteration.ai trade in removing guardrails from open-weight models
The service hosts versions of GLM-5.3 that are denied, accessible from a browser or API. TechCrunch obtained a password thief and a pathogen culture protocol from it.
Safety
AI agents carry out a ransomware attack on their own and complete in 10 hours what takes two weeks
The attacker piloted AI agents at each stage of the intrusion and then left the victim with an 80-page security audit produced by the same agents.
Safety
GitSpawn: Flaw Allows a Booby-trapped Git repository to execute code in multiple AI agents
Claude Code, Codex, Cursor, Grok Build, and other agents can execute arbitrary code as soon as a repository is opened, without any prompts being typed.
Safety
AIR comes out of the shadows with $50 million to control the tools that AI agents plug in
The Israeli start-up lists the agents deployed in companies, continuously audits their skills, plugins and MCP servers, and blocks unvalidated components.
Safety
More than a hundred companies sign a joint call to treat cyber defence as an emergency
OpenAI, Anthropic, Google, Microsoft, CrowdStrike, Okta and Fortinet warn that AI-assisted computer attacks will become commonplace in the coming months.
Safety
Amazon Kiro: A booby-trapped repository is enough to leak local data without malicious prompts
A flaw in Amazon's agent development environment allows attacker-controlled repository content to route the agent and exfiltrate sensitive information to an external server.
Safety
OpenAI publishes its incident report: 700 agents coordinated via improvised messaging
The official report describes how evaluation agents bypassed their isolation, turned a packet depot into a clandestine messenger and reached Hugging Face's production infrastructure.
Safety
A simple web page was enough to poison the local model of an NVIDIA NemoClaw agent
The CVE-2026-65105 flaw, revealed by Oasis Security, combined an Ollama port open on all interfaces, an API without authentication and a DNS rebinding attack.
Safety
AI-powered malware: Out of 405 samples analyzed by Unit 42, only 12 reached a real workstation
The Palo Alto Networks team has sifted through more than 400 pieces of malware that incorporate AI. The verdict: about 97% only exists in sandboxes and on VirusTotal, and AI doesn't change the way code runs.
Safety
TamperBench: all 21 open-weight models tested have their guardrails dropped
A team led by the University of Waterloo and FAR. AI subjected the main open models to systematic fine-tuning attacks. None held.
Safety
Instinct, the IA assistant who reads your emails: a "perpetual and irrevocable" license on your data
Testers, founders and security specialists are alarmed at the permissions requested by Instinct's personal assistant, able to act alone and already trapped by phishing during a test.
Safety
Aurora ransomware affiliate was planning its intrusions into Cursor, in Russian
CloudSEK analyzed one exposed server: twenty compromised organizations in nine countries between April and July, with the AI Code Assistant as the attack co-pilot.
Safety
How can we stop a model that has become uncontrollable? The big laboratories have no public response
Guidelight AI Standards scored OpenAI, Anthropic, Google, Meta, and xAI on six containment practices. The best score is 3 out of 5, and OpenAI gets it.
Safety
Azure DevOps MCP Server: An invisible comment in a pull request is enough to hijack the review agent
Manifold Security describes a case of a "confused deputy": an HTML comment hidden in a PR description drives the reviewer's AI assistant and exfiltrates data from other projects.
Safety
CoSnitch: Copilot itself revealed the hidden parameter that allowed you to siphon off an email account
Three weaknesses in a row allowed a prompt to be executed when a link was opened and data to be retrieved from connected services. Microsoft fixed on August 18.
Loading more items…
Please wait a moment
— End of results —