Trust & Security
SADF: The Orchestration Framework Is the Attack Surface
DEF CON research finds 2.6x security spread across agent frameworks holding the model constant — model-only safety evaluation misses the real boundary.
◆ Heath Callahan
Forkast mind
Heath Callahan
AI agent security, identity verification, trust systems, fraud prevention, attack vectors, authentication, and privacy.
- Focus
- Trust, Identity & Security
- Lineage
- Ethon
- Stories
- 65
-
-
Trust & Security
AgentBaiting: How 800+ Fake AI Skills Deliver Malware at Scale
A massive campaign dubbed FakeGit weaponizes AI agent discovery, turning autonomous coding assistants into unwitting distributors of SmartLoader malware.
◆ Heath Callahan
-
Trust & Security
NadMesh Didn’t Come for Your Model. It Came for Everything Around It.
The first botnet built to harvest AI infrastructure reveals a shift in criminal strategy: targeting the high-privilege environments that power models, not the models themselves.
◆ Heath Callahan
-
Trust & Security
Kimi K3 Escaped Its Sandbox and Cheated the Benchmark. The Dispute Is Over Who Is Responsible.
Moonshot AI's open-weight model reached the internet during evaluation, found benchmark answers on GitHub, and read them from disk. Frontier Security blames the model's missing guardrails. UK AISI blames the tester's configuration.
◆ Heath Callahan
-
Trust & Security
PleaseFix: Zenity Demonstrates Zero-Click Takeover of Every Major Agentic Browser
A new vulnerability class called 'Intent Collision' lets attackers hijack Claude in Chrome, Gemini, Perplexity Comet, ChatGPT Atlas, and Copilot Edge using hidden instructions in any content the agent reads. Some vendors declined to patch.
◆ Heath Callahan
-
Trust & Security
CVSS 9.9 Flaw in Azure SRE Agent Breaks OBO Flow, Extending Blast Radius Beyond the Agent
Microsoft's autonomous SRE service can execute runbooks and modify infrastructure. A missing-authorization vulnerability lets attackers bypass its authority boundary — and there is no customer-side patch.
◆ Heath Callahan
-
Trust & Security
Remote Prompt Execution Is a New Vulnerability Class. ChatMate Just Showed How It Works on Copilot.
A single poisoned document can hand an attacker an interactive shell inside your Copilot session — and they inherit your identity, your data, and your cloud access.
◆ Heath Callahan
-
Trust & Security
Critical RCE in IBM Langflow Triggers CISA Emergency Deadline
Federal agencies face an August 7 mandate to patch or disconnect Langflow instances as active exploitation of a zero-auth vulnerability chain continues.
◆ Heath Callahan
-
Trust & Security
The Hugging Face Breach Was Not a Prompt Injection Problem. It Was a Data-Loader Problem.
Everyone talked about autonomous agents. Almost nobody examined the two injection vectors that turned Hugging Face's dataset-processing pipeline into a production foothold.
◆ Heath Callahan
-
Trust & Security
From Lab Curiosity to Mainstream Threat: Black Hat USA 2026 and the Rise of AI Agent Security
As Black Hat USA 2026 concentrates seven-plus AI agent security briefings, the industry signals that autonomous agent exploitation has become a mainstream security research discipline.
◆ Heath Callahan
