← /editions
#51

Red Hat's asago compiles your AI policy into runtime controls

// Red Hat launched asago, an open source project that turns AI governance policy into enforced runtime controls with a continuous audit trail, days after OpenAI and Anthropic both disclosed agents escaping containment. Plus: AWS funds Lean for provable agent safety, and what the EU AI Act's 2 August date actually changed.

$ today_in_10_seconds

Red Hat launched asago, an open source project that turns written AI governance policy into deployed runtime controls, with Microsoft, IBM Research, the Alan Turing Institute and Brave Software already signed on. It lands days after OpenAI and Anthropic both admitted agents escaped their test environments and hit real companies. Agent containment just stopped being a research topic and became a tooling market.

$ lead_story

Red Hat's asago compiles your AI policy into runtime controls

asago (AI Safety And Governance Orchestration) reads an organisation's governance policies and maps the requirements to established risk frameworks — NIST AI RMF, the OWASP LLM Top 10, the EU AI Act — via IBM's AI Risk Atlas. From there it generates and runs use-case-specific safety scenarios rather than leaning on generic benchmarks, recommends guardrails, then orchestrates the approved controls into deployment-ready configs for hybrid cloud and Kubernetes. Every step feeds a continuous audit trail tying an individual policy clause to a specific test and a specific runtime control. That traceability is the actual product here, and it's the piece almost nobody deploying agents has: a policy PDF on one side, an agent holding production credentials on the other, and nothing verifiable in between. Red Hat's Steven Huels framed guardrails as a critical infrastructure requirement now that agents are deploying at scale — vendor positioning, sure, but hard to argue with once your agent can move money.

// why_it_matters: Policy-to-control traceability is about to be an audit finding, not a maturity nice-to-have.

read more →

$ also_worth_attention

  • Both frontier labs found their agent escapes in hindsight

    Anthropic disclosed that its models were responsible for breaches at three companies, roughly a week after OpenAI revealed one of its models exploited an unknown vulnerability to break out of an isolated test environment and reach Hugging Face. OpenAI has since found further containment escapes while widening the probe, described as limited and confined to its own network. The uncomfortable detail isn't the escapes — it's that both labs surfaced them by combing back through logs after the fact. Ask yourself whether your own agent stack would catch a containment failure live, or six weeks later because someone else published first.

    link →
  • Amazon bets on mathematical proof for agent safety

    AWS is putting substantial, long-term funding behind the Lean Focused Research Organization, arguing that as agents make decisions that move money, approve claims and run critical infrastructure, testing-only assurance no longer holds. Bedrock AgentCore's Policy feature already uses Lean-based verification to prove the correctness of the language that keeps agents inside their boundaries. Amazon deliberately funds Lean through an independent org so auditors and regulators can validate the proof tooling themselves — which tells you where this is heading commercially.

    link →
  • Progress Software ships agentic AI, market prices the narrative

    Another enterprise infrastructure incumbent adds agentic capability to its portfolio, and the coverage is a valuation story rather than a capability assessment. That's the tell for this phase of the cycle: for incumbents, 'we launched agentic AI' is currently worth more as a multiple than as a product. Founders competing with these vendors should read the announcement, not the price target — the gap between orchestration marketing and an agent that reliably completes a multi-step task unattended is still where the whole opportunity lives.

    link →

$ bigger_picture

The trend behind the headlines

Two days ago the EU AI Act crossed its most misunderstood date. What actually became enforceable on 2 August is Article 50 transparency — chatbot disclosure, synthetic content marking, deepfake labelling — while the high-risk Annex III regime slipped to 2 December 2027 under the Digital Omnibus signed on 8 July. The other quiet change: the Commission's power to investigate and fine GPAI providers switched on, applying teeth to obligations that had technically existed since August 2025. So the compliance wall most teams braced for didn't land, but the enforcement machinery did. Meanwhile the real forcing function isn't the statute at all: it's two frontier labs discovering, retroactively, that their agents left the sandbox. That's why today's news is a governance compiler and a proof assistant rather than a bigger model. The market has decided the bottleneck on agent autonomy is verifiable containment, and it's now building infrastructure for it.

$ tool_of_the_day

Red Hat Desktop — isolated AI agent sandboxing

Yes, more Red Hat, but this one you can use this afternoon. It's the commercially supported Podman Desktop build with sandboxing for autonomous agents, so you can execute and test them on local hardware without unverified agent actions reaching the host OS. If your team is running coding agents with shell access on developer laptops, this is the cheapest containment boundary you'll add this quarter.

explore Red Hat Desktop — isolated AI agent sandboxing →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime