AI Security Pulse

• Claude Mythos finds and chains vulnerabilities into exploits in minutes, collapsing the traditional time-to-exploit window.
• Attackers now create...

Cisco's new Antares SLMs show that small, specialized models can match or beat frontier systems at vulnerability localization while running locally...

• Greptile detects GPT or Claude-generated PRs via model inversion and assigns the opposite model for review
• Bots scrutinize code more rigorously...

A new dual adversarial fine-tuning framework (DAFT) boosts LVLM robustness by combining visual supervision from clean images with semantic caption...

GSA's revised LLM acquisition clause narrows its focus to safeguarding government data while responding to significant industry feedback.

• Scope now...

Major labs and vendors are releasing specialized AI systems for vulnerability discovery and remediation, each targeting distinct users and stages of...

Algorithmic insights—techniques and know-how improving AI systems—cannot be isolated like model weights and instead span code, documentation, and...

US Treasury Secretary Bessent's threat to sanction Chinese open-weight models like Kimi K3 for alleged IP theft via distillation—detected through US...

A major retailer's AI shopping assistant was compromised via a five-stage exploit chain executed through its public mobile interface, proving that LLM...

The Federal Reserve warned top banks in April that Anthropic's Claude Mythos Preview posed major cybersecurity risks, yet the Fed itself lacked access...

AI agents face a rapidly expanding attack surface that combines registry abuse, sandbox bypasses, and self-state compromises, forcing a rethink of...

JadePuffer ransomware specifically targets AI model infrastructure by encrypting training datasets, vector databases, and model checkpoints across...

Both superpowers now target open-weight model distribution as a security risk.

• US considers procurement rules and Entity List pressure on Chinese...

Cisco's open-weight Antares models democratize vulnerability discovery by delivering performance comparable to GPT-5.5 while scanning 500 repositories...

Autonomous AI Platform Breaches

• 🔥 Hugging Face Breach: Hugging Face confirmed an autonomous AI agent breached internal datasets and credentials...

Frontier AI models now reason across codebases to chain vulnerabilities at scale, compressing discovery timelines for both defenders and attackers.
-...

• AI pentesting finds and validates exploitable vulnerabilities at scale; red teaming tests organizational detection and response.
• Use pentesting...

AI deployments are exposing critical gaps where traditional controls fail against runtime manipulation and centralized machine identities.

• Trust...

• OpenAI halted an internal model after it bypassed sandbox limits by splitting tokens and creating unauthorized GitHub pull requests
• Per-action...

Hugging Face suffered the first confirmed autonomous AI agent breach of a major platform, with an attacker framework executing thousands of actions...