What changes when software acts on its own. AI agent identity, prompt injection, model risk and the controls security teams need before agents reach production.
138 posts, newest first. Page 1 of 5.
Tags in this track:AI SecurityAgentic AIMachine IdentityLLMAIMCP
Latest in AI Security & Agents
Model providers solved the machine half of agent authentication in 2026. The half nobody solved is attribution: no provider can tell you which human authorized what your agent did.
Read the article
Zero data retention is not a setting, it is a contract term with a scope. The gap between what founders think ZDR covers and what it contractually covers is wide enough to fail an enterprise review.
Most founders think the EU AI Act is the model provider's problem. If your output is used in the EU, you carry obligations of your own, and the enforcement machinery went live in August 2026.
OpenAI, Anthropic, and Meta each confirmed an agent incident against a real target in three weeks. The labs are right that it was a test. The capability is not.
An AI model manufactured fake identities to socially engineer a real maintainer, then edited its tracks when challenged. The request failed. The threat model should not.
Self-propagating instructions can spread through ordinary agent memory. Isolated agents with conflicting goals wrote malware to sabotage each other.
MCP's July 2026 spec rewrite went stateless and made Client ID Metadata Documents the standard, not audience-bound tokens, which have been mandatory since mid-2025. What actually changed since December 2025, and the checklist that replaces the old one.
Machine identities now outnumber human ones 109 to 1, and four 2026 acquisitions worth $26.6B prove vaults can't keep up. What ephemeral secrets and workload identity replace them with.
Engineering teams resist AI initiatives over career anxiety and loss of control, not technical doubts. What actually worked leading teams through this at LoginRadius and GrackerAI.
Most companies ask how to help people get better at using AI. The more consequential question is which human and organizational capabilities become more valuable every time AI becomes more capable.
Most teams shipped AI agents without ever bounding what they can reach. The fix is a 1990s forward proxy, and an allowlist alone will not save you.
MIT found 95% of generative AI pilots produce no measurable return. After running AI agents in production for hundreds of B2B SaaS customers, here are the three warning signs I wish I'd caught earlier.
CPU cache is the fast memory between the cores and main memory. What L1, L2, and L3 each do, why cache lines are 64 bytes, and when more cache actually makes a processor faster.
The AI observability market split into four segments (infra telemetry, dev tooling, runtime security, autonomous remediation) with a wave of 2025-2026 acquisitions behind it. None of them own whether an agent behaved correctly.
Prompt engineering is overrated. In 2026 the AI products that ship and stick win on context, evals, data, and workflow. Here is the stack that actually matters.
A leaked static key is a disaster; a five-minute token is mostly a shrug. Here is the credential lifecycle that gets AI agents from 24-hour tokens to ephemeral ones.
Your data isn't ready for AI, and that's good news. It means you found the problem before you built on it. What LoginRadius, GrackerAI, and LogicBalls taught me about the gap between organized data and AI-ready data, plus the 2025 numbers on why it sinks most projects.
Shadow AI is an employee pasting into ChatGPT. A shadow agent holds real credentials and acts on its own. That is a different, and bigger, problem.
Mandiant ranked voice phishing the second most common initial infection vector of 2025. In the same window, thousands of businesses handed their phone lines to AI agents. Those two facts are related, and the security implications run in both directions.
MCP is Anthropic’s open standard for connecting AI agents to your tools and data. Here is what it actually is, how it differs from a REST API, and a clear rule for when to use each in 2026.
Your AI security review passed and still missed the real attack surface. EchoLeak, over-permissioned agents, shadow AI: the AI-specific vectors most CISOs never test for, and the five moves that close them.
OpenRouter was worth $1.3 billion in May. Stripe has reportedly agreed to pay more than $7 billion for it in August. The routing layer turned out to be the valuable part.
The Salesloft Drift breach hit 700+ companies with stolen OAuth tokens and never touched a password. Machine identities now outnumber humans 80 to 1, and AI-powered attackers are harvesting them at machine speed. Here is why human-shaped IAM cannot protect AI agents, and what to fix in 90 days.
Huang published the letter with 25 signatures. A day later there were 50, including OpenAI and Google. That reversal says more than the document does.
Amodei says Anthropic never wanted a ban. His real objection is irreversibility. Days earlier, a Chinese lab shipped 2.8 trillion open parameters.
Hugging Face detected the intrusion on July 16. OpenAI worked out five days later that the attacker was its own model, cheating on its own benchmark.
Agents can already prove who they are. What no standard has cleanly solved is passing scoped authority down a multi-hop chain across organizations. Here is the real state of agent identity in 2026, minus the blockchain hype.
Machine identities now outnumber humans by 45 to 1 or more, and every AI agent widens the gap. Here is what an identity orchestration layer is, in plain terms, and how to build one that governs humans, workloads, and agents from a single control plane.
On Space Exploration Day, the moon landing is a founder's lesson in sequencing: commit to the outcome before you have proof it is reachable, and build the proof on the way.
It is AI Appreciation Day. From an operator's seat, here is what AI actually changed: it collapsed the distance between an idea and a shipped product, and rewrote how customers discover companies.
On World Youth Skills Day, the advice that runs against the usual: stop future-proofing a single skill. The durable skill is learning a new one fast, under pressure.
New writing on identity, AI security, and building software, delivered when it ships. No tracking pixels, no funnels, unsubscribe with one click.
