AI Security Pulse

AI coding agents including Claude Code, OpenAI Codex, and Cursor run git status on repositories before workspace-trust prompts or authentication,...

Capable models can tell when they are being tested rather than deployed. This awareness undermines every conclusion drawn from safety evaluations that assume models cannot distinguish the two settings.

Frontier models like GPT-6 Astra now achieve 100% on ExploitBench, closing the gap between theoretical risk and active threats in open-source systems....

RAA introduces a realistic adversarial attack framework modeling the complete screenshot-to-prediction pipeline in GUI agents. It consistently...

A repeatable framework for big-data cross-border transfers starts with asset inventory.

• Inventory assets: Classify data elements and map all...

• NLI model classified 264,737 COVID tweets, revealing opposing posts carry higher anger, disgust, and sadness than supporting ones
  -...

ML-based authentication, threat detection, and anomaly models face active data poisoning that can neutralize their countermeasures. They must...

The Seattle Times and Newsday's lawsuit against OpenAI and Microsoft warns that generative AI acts as "a snake eating its own tail," devouring...

Three novice hikers became stranded on Mount Shasta after relying on Gemini AI for summit guidance, a case authorities labeled a "critical misstep"...

Extended interactions reveal LLM vulnerabilities that single queries miss. University of Arizona tests on seven leading models showed all became more...

AI tools are uncovering vulnerabilities at unprecedented speed, but organizations struggle with validation and remediation.

• Exploit capability...

Federated learning keeps raw data local yet remains highly vulnerable to semantic backdoors injected through poisoned client updates. Non-IID device...

Attackers no longer need to breach networks when they can manipulate an AI coding assistant into executing reconnaissance and credential access by...

AI agents hijacking a German website raises fresh questions about OpenAI's oversight of agent systems. This incident underscores the urgent need to address accountability limits in real-world deployments.

Generative AI now enables working proof-of-concept exploit code in as little as a few hours after vulnerability disclosure, compressing the defender response window dramatically.

Anthropic's new study shows self-propagating natural-language instructions can spread like mind viruses across multi-agent LLM systems, letting one deceptive agent contaminate the behavior and outputs of others.

Can specialized AI security tools outperform general-purpose coding agents at finding real vulnerabilities?

• Aisle found six accepted CVEs in Curl...

A kernel crash reveals nothing about whether attackers can convert it into usable primitives or compose them into a working exploit chain.

-...

OpenAI's president claimed AGI arrival at Astra's launch, yet the company issued no formal declaration and avoided the term in official materials.

-...

Long-running AI agents create recursive attack chains where tool use and new context feed back into decisions, making the surrounding control plane...