Scribbled thoughts and views on AI technology, security, and its impact on culture, from the team at raxIT Labs.
GPT-6 Astra and Claude Fable 5.1 shipped this month. If you plan to run them in production, some of the risk in their own system cards is yours to accept rather than mitigate. This is a walk through which parts.
A deterministic decision point outside the model, in the live path, that decides before every action executes: that's the TLDR of every mitigation in the 2026 OWASP LLM Top 10. Excessive Agency jumped from rank 6 to rank 3, checked for the first time against 7,714 real incidents instead of a practitioner vote alone. Read the mitigations, not the ranks, and they all converge on one prescription: mediate everything.
The 15 July framework binds data centres: where they sit, what power and water they draw, what they pay for the grid, plus training on Australian creative work. It puts no new obligation on anyone building with models. What already binds you is the law you have: consumer law, APRA, ASIC, the Privacy Act. Neither document says who gets to say no.
Claude Tag puts @Claude in Slack as its own per-channel service account, not on behalf of the person who asked. That breaks the delegation model everyone converged on, where the audit log can still name a human. Useful, but it revives the old service-account problems with an AI on top.
A God Agent is one model wired to every tool on one shared identity, holding the whole lethal trifecta at once: it reads untrusted content, touches private data, and talks to the outside world. You can't filter prompt injection out of that, so you change the architecture instead. Three deterministic moves: scope every agent, sign every call, stop every breach. Break even one leg of the trifecta and the blast radius shrinks.
Agent identity makes more sense as four stacked layers: a tamper-proof token format, cryptographic proof of which workload is running, a delegation chain that keeps the human as the subject, and a way to onboard to a service an agent has never met. Climb all four and you've proven who the agent is. You still haven't proven that what it did was okay.
Misalignment maps onto vulnerability classes security engineers already operate on: backdoors, defense evasion, privilege escalation, exfiltration. Calling it ethics keeps it off security teams' desks. Reframing it as security decides who owns the work, which budget pays, and which playbook applies.
Claude 4.7 doesn't rely on one safety mechanism. It stacks a rulebook, trained refusals, differential capability reduction, two runtime probes, and a live feedback loop. Understanding which layer blocks what matters if you're building on the API.
BodySnatcher (CVE-2025-12420) showed how AI agents with aggregated permissions can compromise entire platforms in seconds. Traditional security controls designed for humans don't work at machine speed. Organizations need threat modeling and runtime controls for all three layers: API auth, identity binding, and agent execution.
The EU (9/10 risk), US (5/10), and Australia (6/10) take vastly different approaches to AI regulation. Build for EU standards globally: the Brussels Effect means you'll need them anyway.
AI agents are breaking traditional identity and access management systems. From impersonation risks to cross-domain delegation chains, enterprises need new frameworks that balance autonomous operation with accountability and security.
When AI appears to think but actually pattern-matches toward desired outcomes, you get sophisticated-looking failure. This fictional crisis demonstrates real research about AI limitations and how to build better systems.
Shadow coding—developers using unauthorized AI tools or code—is creating significant security and compliance risks. Organizations need balanced governance that enables innovation while maintaining security.
Claude 4 models introduce novel enterprise considerations including high-agency behaviors, self-preservation instincts, and potential consciousness indicators that may require enhanced risk management depending on your deployment context.
Learn key AI safety and security principles by examining the detailed instructions within a publicly available system prompt, showing how LLMs can be guided towards responsible behavior.
We translate academic benchmarks into actionable risk signals through our proprietary AI governance pipeline, helping organizations implement regulatory-ready controls.
Organizations need AI governance and security now more than ever. We're building a platform to make powerful protection accessible to everyone.
