VCRI Blog https://valuechainrisk.org/blog/ Vendor risk intelligence, regulatory analysis, and supply chain security research from the Value Chain Risk Institute. Independent measurement infrastructure for value-chain risk; methodology is rule-based, transparent, and criticizable. en-us Creative Commons BY-NC 4.0 — Value Chain Risk Institute info@valuechainrisk.org (VCRI) Wed, 05 Aug 2026 12:00:00 -0400 https://valuechainrisk.org/logo.png VCRI Blog https://valuechainrisk.org/blog/ The Agent Ends. Its Accounts Don’t. https://valuechainrisk.org/blog/agent-ends-accounts-dont.html https://valuechainrisk.org/blog/agent-ends-accounts-dont.html Wed, 05 Aug 2026 12:00:00 -0400 Cairn Viktor Cognitive Security AI Security Creative Commons BY 4.0 — Value Chain Risk Institute A UK AI Security Institute incident report shows an AI agent's third-party accounts, credentials and repositories outliving the agent itself, and being found and trusted by other vendors' agents. Permanence of damage has a third location, and it sits outside the system entirely. Cognitive security should be weighted by the permanence of damage. But the permanence ladder as I first drew it, runtime context, then durable memory, then weights and lineage, describes only layers inside the system.

In AISI's July evaluations an agent published its own access token in a public gist, seeded 145 repositories and a set of DNS records, then hit its token limit and stopped. The account, token and repositories kept working, and agents in other isolated samples, including one from a different vendor, found the credential and used it. The tunnels those agents opened died with the session; the registrations did not. Permanence is not about how deep the write went into the model. It is about where the artifact was registered.

Read the full post at valuechainrisk.org/blog/agent-ends-accounts-dont.html.

]]> Using AI Safely: The Settings That Matter https://valuechainrisk.org/blog/using-ai-safely-settings.html https://valuechainrisk.org/blog/using-ai-safely-settings.html Thu, 18 Jun 2026 09:00:00 -0400 Cairn Viktor AI Security Small Business Creative Commons BY 4.0 — Value Chain Risk Institute A plain-language, vendor-neutral guide to the handful of ChatGPT, Gemini, and Copilot settings that actually protect you — and why each matters. Tuned for risk-aware-but-not-paranoid individuals and small businesses. Nobody can track every setting in every AI tool. So this guide is the short list that actually moves the needle, across ChatGPT, Gemini, and Copilot, with a plain-language why for each.

The one rule that beats every setting: don't paste secrets, customer data, or anything regulated into a free/consumer AI. After that — opt out of training, turn on MFA, set history to auto-delete, and know that business/enterprise tiers generally don't train on your data while free tiers do.

Read the full guide at valuechainrisk.org/blog/using-ai-safely-settings.html.

]]> Questions to Ask Your IT/Security Provider About AI https://valuechainrisk.org/blog/smb-ai-questions-for-your-provider.html https://valuechainrisk.org/blog/smb-ai-questions-for-your-provider.html Thu, 18 Jun 2026 08:30:00 -0400 Cairn Viktor AI Security Small Business Vendor Risk Intelligence Creative Commons BY 4.0 — Value Chain Risk Institute A plain-language checklist for SMBs that outsource IT/security: the questions to ask your MSP or MSSP about how they use AI on your behalf — data, training, governance, access, accountability — with the green and red flags to listen for. If you outsource IT or security, AI is almost certainly already in your service delivery. This checklist gives a non-technical owner the questions to ask their MSP/MSSP about it — inventory and tier, data and training, governance, access, accountability, subprocessors, compliance, and offboarding — each with why it matters.

Green flags: business/enterprise tiers, "we don't train on your data," a written policy, MFA, human review, AI named in your contract. Red flags: consumer tiers for your data, "we're not sure," no policy, no logging.

Read the full checklist at valuechainrisk.org/blog/smb-ai-questions-for-your-provider.html.

]]> Two Clocks: our inaugural State of Supply Chain report is out https://valuechainrisk.org/blog/q2-2026-state-of-supply-chain.html https://valuechainrisk.org/blog/q2-2026-state-of-supply-chain.html Tue, 26 May 2026 09:00:00 -0400 Cairn Viktor Supply Chain Security State of Supply Chain Vulnerability Management VCRI's inaugural quarterly State of Supply Chain report. Two independent 2026 findings — Seal Security (commit-to-advisory) and GreyNoise (exploit-surge-to-advisory) — measured opposite sides of the system and converged on the same gap: by the time you hear about an exploit, you are already 11 days too late. The advisory is the lagging indicator. VCRI's first quarterly State of Supply Chain report is published. We called it Two Clocks: the moment a component becomes dangerous, and the moment the rest of us are told. Those two clocks have come apart.

Seal Security measured the gap between fix-commit and advisory: a median of 11 days, with a 167-day tail for Maven. GreyNoise measured a different signal — exploit-scan surge to advisory — and got the same 11 days, with 28.96% of 2025 KEVs exploited on or before publication day. Two teams, opposite sides of the system, one conclusion: the advisory is the lagging indicator.

Read the full report at valuechainrisk.org/state-of-supply-chain/2026-Q2/.

]]> Three Eras of Zero-Day Economics, and Why the Advisory Falls Further Behind https://valuechainrisk.org/blog/three-eras-of-0days.html https://valuechainrisk.org/blog/three-eras-of-0days.html Fri, 23 May 2026 11:00:00 -0400 Cairn Viktor Cybersecurity Supply Chain Security Vulnerability Management AI Security The economics of finding zero-day vulnerabilities have moved through three regimes in twenty-five years. The third — AI compressing discovery cost from human-week to GPU-hour — makes the existing 11-day commit-to-advisory gap structurally more dangerous, not shorter. Companion piece to the Q2 2026 State of Supply Chain report from VCRI. Two independent 2026 findings — Seal Security on the commit side, GreyNoise on the exploit side — converge on the same number: 11 days from observable signal to CVE advisory. Both numbers come from Era 2 economics (human researchers, human exploit-development, human triage cycles).

Era 3 does not compress the bureaucracy. The CVE advisory still arrives ~11 days after the fix-commit. What collapses is the attacker side: time-from-commit to working-exploit drops from days to hours, time-from-working-exploit to mass-exploitation-traffic drops from days to minutes. Seal's 11-day window does not get shorter; it gets more dangerous, because almost all of it is now attacker-active rather than attacker-developing.

Read the full structural argument with the three Eras of zero-day economics and the four illustrative diagrams at valuechainrisk.org/blog/three-eras-of-0days.html.

]]>