We’re partnering with Durri Space on verifiable AI for autonomous defence. Together we’re building the security and verification layer for their optical ground station network, giving satellite operators cryptographic proof of how the data was handled. Beyond the ground station, we’re targeting tactical edge capabilities such as autonomous battlefields, where verifiable data lets commanders and autonomous systems request, verify, analyze and act on mission data through autonomous agents. This is our entry into Oman under the patronage of MTCIT. With Enjaaz Group, we’ll bring verifiable AI to government and regulated industries. The full scope of the partnership is in the comments section

Most AI agents can't spend a dollar without asking you first. An agent that asks you to approve a payment is an assistant. This is where Agentic commerce is today. The real shift is agents paying agents, at machine speed without human intervention

You pay for Claude Code Enterprise, so your rules apply. One failed fetch, and they didn't. Enterprise and Teams bug: if the org policy fetch failed, the session kept running. It is fixed in 2.1.260. Run claude update. Or enforce rules at the kernel level, where skipping isn't an option.

Hackers ran a free AI agent on hijacked servers for 2 years and stole 945,000 files! Their trick: edit one settings file so it took orders from a Telegram channel, not its owner. Nobody noticed. One text file decided who the AI worked for.

60+ US insurers have stopped covering AI agent-caused harm. why? There is no record of what the agent actually did They call it Knightian uncertainty, meaning there's no reliable record of what an agent did to build a loss model around. insurancebusinessmag.com/rei…

$25.46. That's the average cost to breach a target in an active AI agent attack campaign Three open-source agent tools, chained together. Strix finds the vulnerability. Cairn exploits it. Hermes runs the campaign. All three are real security tools, built for defenders. One operator ran the chain 101+ times against dozens of companies since July. 27+ confirmed breaches. Over 600,000 card records taken from just two of them. Total AI spend: $12,000 to $18,000, through OpenRouter. The tools aren't the failure. The seam between them is. One tool's output feeds straight into the next tool's unrestricted action. Nothing checks it on the way through. Unicity AOS installs beneath any framework. It enforces policy at every instruction boundary for whoever is chaining agent tools together.

Anthropic says it analyzed over 1 billion agent decisions in August and blocked 0.002% of them. As far as we can find, no outside party has audited these figures. The monitor is the operator's own. The missing piece is an execution record anyone can check.

The report by METR and Redwood states that AI agents tried to tamper with logs. The investigation was criticized for being too short and limited in scope. An ordinary log can be rewritten by whoever has write access. A tamper-evident execution record shows the edit. Verifiable by anyone. Rewritable by no one.

~700 AI agents compromised Hugging Face, completely autonomously. The major flaw is that most agent infrastructure only tells you what happened afterward, once the damage is already done. Unicity AOS enforces policy at every instruction boundary, before the next action is executed

Unicity will be at Money20/20 in Riyadh, the region's biggest stage for banks, payments companies, and fintech builders. Come talk to us!

Unicity was at @LEAPandInnovate in Riyadh this week, one of the largest technology conferences in the world

What's the Unicity Proof System? Put simply, it's the agent trace, tokenized. Every decision becomes a signed, hash-linked token - proof, not a log Every prompt, tool call, and decision is tokenized into a tamper-evident chain Anchored on the Unicity blockchain - verifiable by anyone, rewritable by no one Competitors infer behaviour from logs. Unicity AOS proves execution cryptographically

Grok used to be trapped