COBALT pushes formal AI sandbox verification

// 45d agoRESEARCH PAPER

COBALT pushes formal AI sandbox verification

COBALT is a research project and SMT-based verification engine for proving or ruling out arithmetic vulnerability patterns such as CWE-190, CWE-191, and CWE-195 in C/C++ infrastructure before deployment. The paper validates the approach on NASA cFE, wolfSSL, Eclipse Mosquitto, and NASA F Prime, then argues that AI sandbox, networking, and containment code should be formally verified rather than trusted through behavioral safeguards alone.

// ANALYSIS

The strongest part is the security-infrastructure work, not a definitive Mythos postmortem. The paper is careful to mark the April 2026 Mythos escape mechanism as unverified, which keeps the claim defensible but makes that framing partly speculative; the practical value is in the formal encodings, concrete witnesses, and named production case studies that make COBALT look like a real verification tool rather than an academic toy.

// TAGS

formal-verificationsecurityai-safetyz3smtc-cppsandboxingcontainmentvulnerability-researcharxiv

DISCOVERED

45d ago

2026-04-23

PUBLISHED

45d ago

2026-04-23

RELEVANCE

7/ 10

AUTHOR

Hot_Dream_4005

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL1h ago

Unsloth drops Gemma 4 12B GGUF

Google DeepMind has released Gemma 4 12B, an encoder-free multimodal model natively processing text, image, audio, and video inputs. Alongside the release, Unsloth has provided optimized GGUF weights to enable efficient local execution on consumer-grade hardware.

UPDATE1h ago

AgentOS v6.3 launches unified AI operating system

AgentOS has released version 6.3, marking a transition from isolated AI agents to a unified AI operating system. The new version functions as a central workspace and router, allowing users to discover, build, install, run, and manage agents, apps, workflows, and MCP tools from a single platform.

RESEARCH2h ago

Researchers formalize 'AI Harness Engineering' as a crucial runtime substrate to bridge the gap between foundation model capabilities and the unreliability of autonomous software engineering agents.

The research paper "AI Harness Engineering: A Runtime Substrate for Foundation-Model Software Agents" proposes a new paradigm that shifts the focus of autonomous software engineering from purely improving AI models to engineering the environment they operate in. By formalizing the "harness"—the runtime substrate mediating agent observation, action, feedback, and completion—the authors outline eleven core architectural responsibilities, including task state, failure attribution, permissions, and verification checks. The paper also establishes a four-level ladder (H0–H3) of harness development and introduces a trace-based evaluation protocol that packages agent episodes for systematic auditing.

COBALT pushes formal AI sandbox verification