< ciso
brief />
Tag Banner

All news with #ai red teaming tag

161 articles

Bigger Prompt Injection Datasets Can Weaken AI Agents

🛡️ Researchers from the University of Waterloo, Zhejiang University and KTH Royal Institute of Technology found that AI agents retrained on larger piles of prompt injection attacks were easier to hijack in three of four tests. The likely culprit is that automated red-team tools keep rewording the same few tricks. Their PIPETTE system rewards technique diversity and cut off-the-shelf attack success to under half a percent on AgentDojo.
read more →

Google Warns of Imminent Fully Agentic AI Attacks

🔍 Google’s Threat Intelligence Group warns that fully autonomous, agentic attacks could emerge within three to six months after observing a shift toward pen-testing frameworks like ARTEX. The GTIG AI Threat Tracker cites increased use of open-source multi-agent frameworks such as CAIRN, CyberStrikeAI and PentAGI, and notes actors are experimenting with custom harnesses. These tools orchestrate sub-agents across multiple kill-chain stages and use LLMs for on-demand payloads, increasing speed and scale while currently trading stealth for velocity.
read more →

FBI warns of growing abuse of AI pentesting tool

🛡️ The FBI has warned that the open-source AI penetration-testing framework ARTEX is being abused by malicious actors worldwide, with campaigns observed since at least August 2026. Targets span industries including finance, government, technology, defense, healthcare and telecommunications, and incidents have caused large-scale data exfiltration and supply chain compromises. The Bureau highlighted that ARTEX integrates with external LLMs, accelerates attack chaining to under three hours in some cases, and lowers the barrier to sophisticated intrusions. It urged organizations to adopt recommended security controls and share indicators related to ARTEX activity.
read more →

Adapting Defenses to AI Penetration Orchestrators

🛡️ Over the past month, open-source LLM-based penetration testing and orchestration frameworks have proliferated rapidly, with ARTEX and its many forks spawning hundreds of similar projects that have scanned Cloudflare customers worldwide. These autonomous tools can map, enumerate, and exploit environments at machine speed, chaining stolen credentials and enabling rapid lateral movement. Cloudflare recommends a six-pillar Modern Security Model to contend with the new volumetric and automated threat.
read more →

Anthropic Shares Guidance for Building Strong AI Safety Evals

🛡️ Anthropic outlines how to build more reliable AI safety evaluations, arguing that a safety claim is only as sound as the eval behind it. Drawing on reforms from psychology and its own system card work, the post stresses expert teams, sandboxed testing, and designs matched to the question being asked. It also recommends measuring appropriate responses rather than simple refusals, and using multi-turn simulations to capture realistic risk.
read more →

AI-assisted discovery of long-lived Linux kernel bugs

🔎 A Microsoft FORGE Lab project named MDASH combined AI agents and human researchers to scan selected Linux kernel subsystems, generating 26,502 candidate findings in about two weeks and narrowing them to 59 public upstream reports. AI-produced proof-of-concept tests and automated screening produced 621 promising cases, with human experts validating root causes, drafting patches, and coordinating with maintainers. Of the 59 reports, 38 fixes were merged into mainline and eight had working local privilege-escalation exploits; many bugs had persisted for years.
read more →

AI-Enabled Red Team Tools Accelerate Attacker Pace

🔍 Unit 42 reports that AI-enabled red-teaming platforms like ARTEX and CyberStrikeAI are being adopted by threat actors to perform rapid, autonomous intrusion campaigns. Observed activity includes internet-facing consoles, coordinated cloud-targeting, and forks filling gaps after ARTEX was closed-source. The report outlines tool categories, behavioral hallmarks, and practical mitigations defenders can apply.
read more →

Lessons from 59 Linux kernel bug reports

🛡️ At Microsoft FORGE Lab, codename MDASH scanned selected Linux kernel subsystems from late May to mid‑June 2026, producing 26,502 raw findings that were triaged into 621 PoC-backed cases, 59 upstream reports, and 33 CVEs. Human experts validated and worked with maintainers to merge 38 fixes into mainline; eight CVEs had confirmed local root exploits. The post reflects on scaling discovery with AI and the continued need for expert review.
read more →

Anthropic Glasswing: Findings Overwhelming Maintainers

🔍 A VulnCheck report by Patrick Garrity critiques Anthropic’s Project Glasswing, finding a large volume of AI-flagged issues but few verified and fixed vulnerabilities. Anthropic’s dashboard lists 37,966 candidate findings with 5,276 validated unique vulnerabilities, yet only 217 have been fixed. Anthropic says triage capacity is a bottleneck and has launched an OSS Scanner to let maintainers self-triage, a move that has shifted significant workload to open source maintainers.
read more →

Microsoft finds long‑standing Rust standard library bug

🛡️ Microsoft researchers used their AI-powered Hyenas system to uncover multiple memory-safety and undefined-behavior bugs in Rust’s standard library, including a nearly 11-year-old issue in process::exit. Two of the findings affect stable Rust and could lead to use-after-free or incorrect signal handling under unusual conditions. The team reproduced issues, reported fixes to Rust maintainers, and assessed practical exploitation as unlikely.
read more →

FORGE Lab finds memory-safety issues in Rust std

🔍 Microsoft FORGE Lab used its AI-assisted Hyenas scanner to find memory-safety bugs in Rust's std and core libraries, including cases where safe Rust can trigger undefined behavior. The research highlights three recurring patterns of misplaced trust in unsafe code: assuming code won't panic, trusting user-provided implementations, and trusting the underlying system. Researchers reported multiple issues to Rust maintainers and discussed fixes and mitigations.
read more →

OpenAI unveils GPT-6 Cyber with stronger security

🔒 OpenAI has released GPT-6 Cyber, a security-focused model designed to better find and fix vulnerabilities while remaining aligned to researcher-defined boundaries. The company said the model classifies as Critical under its Preparedness Framework and demonstrates improved adherence to task limits versus GPT-5.6 Cyber. Early results include higher ExploitGym performance and discoveries of numerous Linux privilege escalations and high-severity browser vulnerabilities. Access is initially restricted to select Daybreak Red customers via API and Codex while additional guardrails and misuse detection are enforced.
read more →

OpenAI’s GPT‑6 Cyber Elevates Vulnerability Discovery

🔍 OpenAI released GPT-6 Cyber, a restricted-access model designed to help security researchers identify, validate, and prioritize exploitable vulnerabilities and harden patches. Researchers report using the model to find 81 previously unknown high-severity Firefox bugs and to develop exploits for hundreds of Linux kernel flaws, including many novel techniques. OpenAI emphasizes stronger alignment and safeguards, limited Daybreak Red access, and engagement with vendors to coordinate fixes.
read more →

Joe Security launches Adaptive Detonation for sandboxes

🛡️ Joe Security introduces Adaptive Detonation, an AI-driven capability for Joe Sandbox that actively intervenes when malware samples remain inert during initial sandbox runs. The system uses AI agents to analyze execution traces, identify blockers such as timing checks or missing dependencies, and apply targeted interventions like creating files, adjusting environment settings, or supplying command-line arguments. After modification, the sample is reanalyzed to confirm whether the intervention reveals concealed malicious behavior.
read more →

OpenAI Pipeline Produces 948 Linux Privilege Escalations

🔍 OpenAI detailed a five-stage agentic pipeline that used a cyber-focused AI model to find, validate, and patch Linux vulnerabilities. In a mid-October 2026 run the system reproduced 1,724 bugs and generated 948 working local privilege escalation (LPE) exploits and 291 information-leak PoCs for a common Linux distribution, costing $1.2M in API credits. The research highlights both accelerated offensive capability and automated patching potential, while urging upgraded organizational patching and testing processes.
read more →

Introducing Physical Engineering’s Last Exam Benchmark

🛠️ Today Google AI & Infrastructure and Mercor introduce the Physical Engineers' Last Exam (PLE), a new benchmark and open competition to test agents on mechanical, electrical, civil, and structural engineering tasks. PLE provides 29 public tasks, grading tools, and a leaderboard at ple-bench.com, and includes long-horizon, real-world project folders with diverse file types. The benchmark evaluates code compliance, safety, and cost efficiency, revealing a wide gap between frontier models and professional engineering requirements.
read more →

Armadin report: most severe exposures bypass advisories

🔍 New Armadin data shows 88% of critical and high-severity weaknesses discovered during AI-driven live attack tests this year would not have been detected via vendor advisories or external sources. The findings, published November 2 by Armadin, indicate most severe issues stem from customer-written software and misconfigured vendor products, with many attack chains beginning without credentials. Armadin urges stronger controls over service account credentials and pipeline checks.
read more →

OpenAI Daybreak Red now in AWS GovCloud (US‑East)

🔒 Daybreak Red from OpenAI is now available to eligible customers on Amazon Bedrock in AWS GovCloud (US‑East), providing an AI model tailored for authorized vulnerability research, exploit reproduction, and mitigation development. Built as part of the Daybreak cyber defense initiative, Daybreak Red: GPT‑5.6 Cyber includes stronger identity verification, monitoring, and access controls, and runs on Bedrock’s next‑generation inference engine with zero‑operator access at the chip. Inference data is not used to train models and Daybreak Red data need not be shared with OpenAI. Access requires enrollment through OpenAI and coordination with your AWS account team.
read more →

Agentic Offensive Frameworks Accelerate Intrusion Pace

🛡️ Since late summer 2026, Google Threat Intelligence Group (GTIG) and Mandiant have observed threat actors leveraging open-source agentic pentesting frameworks such as ARTEX, PentAGI, Hexstrike, and CAIRN to automate broad portions of the attack lifecycle. These multi-agent, LLM-driven tools rapidly chain reconnaissance, lateral movement, privilege escalation, and exfiltration steps, enabling intrusions at machine speed. GTIG and partners outline observed tradecraft, examples of exploitation, and recommend defensive best practices including control-plane isolation, short-lived credentials, least privilege, segmentation, and behavioral detection.
read more →

AI-Driven Red Teaming Reshapes Offensive Security

🛡️ Over the past 18 months, AI has fundamentally changed offensive security economics, enabling small, AI-equipped teams to outproduce larger traditional red teams. The article distinguishes between AI red teaming (testing LLMs) and red teaming with AI (using AI as an operator’s tool), arguing the latter is transforming operations across reconnaissance, attack, and reporting. Organizations must budget for frontier models, adopt continuous adversarial testing, and hire practitioners who pair deep judgment with AI fluency.
read more →