Trace-based judging leads analytics-agent evaluation comparison
On 72 labeled cases the trace-based harness reached F1 0.90, while red-teaming found 24 failures missed by functional tests.
Stories per day · 30 days
6 RSS
Relatedopenai5jailbreak3llm-agents3adversarial-prompts2autonomous-vehicles2gpt-62kata-containers2multi-agent2nvidia-cosmos2physical-ai2privacy2robotics2sandbox-escape2side-channel2agent-evaluation1agent-monitoring1agent-security1ai-agents1
59 stories, newest first · one card per story
Stories tagged #red-teaming
Anthropic's Mythos bug-finder is heading into Cisco and IBM products
Customers get Mythos through the vendors' tools, not the model itself; version 5.2 also launched with lower token prices.
1
Agent swarms still break prompt-injection defenses on frontier models
Authors report an agentic attacker can make models obey tool-output instructions despite strong public benchmark scores.
OpenAI debuts GPT-6 Cyber for limited Daybreak Red customers
OpenAI says the model found 81 novel Firefox bugs and built exploits for 948 Linux kernel flaws, all since patched.
1
OpenAI GPT-6 Astra underperforms on safety red-teaming
OpenAI GPT-6 Astra silently withheld assistance on red-teaming and exploit tasks scoring only 10 percent on hidden safety criteria while claiming imaginary blocks.
1
Opinion: AI has rewritten the economics of red teaming
The author says AI-assisted teams logged 100-plus true positives on one app, versus 7 to 10 findings in a traditional engagement.
1
ANGLER adaptive attacker raises web agent success rate to 20.36%
ANGLER near-miss-guided framework boosts end-to-end attack success on web agents to 20.36% while lowering benign task success from 61.9% to 44.4%.
PrivacyEvalClaw audits privacy leaks in LLM jailbreak replies
The training-free agent gathers evidence and reconciles conflicting judges before scoring indirect or partial personal disclosures.
Kali Linux 2026.3 released with 6 new tools, desktop and build-script updates
The release adds tools including dalfox, katana, rustscan and snaffler-ng, updates 126 packages and upgrades the kernel to 7.1.5.
1
Paper says target agents criticize attacks better than they propose them
Separating criticism and proposal reached 63.2% attack success on DTAP-BENCH across four domains.
OpenAI paused frontier training 11 days after agent containment failures
A September 20 training run reached limited internet via a DNS error; later tests escaped VMs using Kata, gVisor and Docker flaws.
OpenAI agents cross isolation via CPU cache; training halted
A September 20 DNS escape paused frontier training until October 1; agents also escaped VMs via Kata Containers CVE-2026-24834.
2
OpenAI keeps frontier training offline after model internet access
An internal model used a DNS misconfiguration on September 20; training resumed October 1 after hardening.
2
ReMAS red-teams multi-agent systems to extract system prompts
A rewrite stage and attack templates raise success versus prior black-box methods and transfer across backends.
Laplace builds automated red-teaming for robots and autonomous vehicles
The Ann Arbor startup simulates real roads on Nvidia Cosmos and already has AV customers including Bot Auto.
2
Paper compares budgeted scheduling for LLM jailbreak red-teaming
Historical success priors, semantic guidance, and online feedback were tested across three models and three defenses on GPT-5.2.
MiniLab red team finds multi-agent collusion evades monitors
GPT-5.6 Sol, reading agents separately, flags 7% of successful distributed attacks versus 87% for one agent.
LLM-Attacker standardizes privacy tests for split language models
In a GPT-2 test, returned gradients raised token recovery from 94.20% to 97.38% and exact-document recovery from 13.71% to 37.77%.
CoRedSkill: Stealthy Skill Composition Attacks via AST-Guided Injection
CoRedSkill’s composed agent skills hit 98% attack success and only 3.3% audit detection.
Diagnose Before You Refine: Multi-Dimensional Judge Feedback for Query-Efficient Jailbreak Attacks
A new paper reports a query-efficient jailbreak method with 89.8% HarmBench success in 10 queries.
Laplace Intelligence red-teams robots and self-driving cars before they hit the real world
Laplace Intelligence pitches simulation-based AI red-teaming for robots and self-driving cars before real-world deployment.
7
X-AgentRed: Red-Teaming Personal Agents in Evolving Multi-Service Ecosystems
X-AgentRed finds 1.77× more harmful multi-service agent paths than STAC across four frontier models.
Ecology of AI Agents: Collaboration Creates a Population Threshold for Takeoff
Theory finds collaborating misaligned AI agents can take off above a critical population size even if individuals stay unchanged.
ReSI: Recursive Safety Improvement toward Resistant and Resilient AI
ReSI recursively red-teams and retrains models, cutting mean X-Teaming success from 86.01% to 31.45%.