On 7 April 2026 NIST issued a concept note for an AI risk profile aimed at critical infrastructure. The base framework stays voluntary. Operators still need practices for the system they actually run, not a demo.
nist.gov/itl/ai-risk-managem…
9
Privacy programs love counting controls. People experience outcomes. If a compliant system can still expose, exclude, or manipulate someone, the risk model is measuring the wrong thing.
doi.org/10.6028/nist.cswp.40
4
The EU terms-and-conditions database holds over 790 document sets from more than 400 services and reviews them several times a day. Policy text that changes in the dark is not a contract the user can follow.
digital-strategy.ec.europa.e…
3
If AI could do most coding, 75 percent of developers would still ask a person when they do not trust the answer. The survey puts humans in the role of final check. A workflow that removes that check has left the evidence behind.
survey.stackoverflow.co/2025…
6
The worst AI code is not obviously broken. It is almost right, passes the happy path, and leaves the expensive debugging to a human.
Should AI-generated changes face stricter review than human changes?
survey.stackoverflow.co/2025…
4
52 percent of developers either skip AI agents or stay with simpler tools, and 38 percent have no plan to adopt agents. The majority is not waiting for a better prompt. They are refusing unattended change.
survey.stackoverflow.co/2025…
2
General-purpose AI providers must document capabilities, limits, and training-content summaries in the EU. That is not anti-innovation. It is the minimum evidence required to sell power responsibly.
digital-strategy.ec.europa.e…
15
The Commission says a minor change to a general-purpose model does not make you its provider. A significant modification does. The guideline is not a statute, but it is what enforcement will follow.
Who draws that line in your release process?
digital-strategy.ec.europa.e…
1
1
44
MFA, logging, and SSO should not be premium security features. CISA says secure products should include them without extra cost. Charging for basic defense turns risk into an upsell.
cisa.gov/securebydesign
5
Shipping critical-infrastructure software that already contains a known exploited vulnerability is a bad practice. CISA says patch those components before release. A launch that includes an active flaw is an incident with a marketing date.
cisa.gov/resources-tools/res…
1
7
Experienced developers trust AI less, not more. That may be wisdom, or it may be a profession defending its status. What evidence would actually change your mind about agent-written production code?
survey.stackoverflow.co/2025…
2
Hardcoded credentials in source are a product defect, not a developer habit. CISA tells manufacturers to pull secrets from a manager in the pipeline and to scan for them before release. If the key is in the repo, the customer cannot rotate it.
cisa.gov/resources-tools/res…
8
On 26 July 2024 NIST published a generative AI profile of its risk framework, NIST-AI-600-1. It is voluntary. A profile that names unique generative risks still leaves the vendor free to ignore it. Voluntary is not a control.
nist.gov/itl/ai-risk-managem…
4
CISA says logs of configuration changes, sign-ins, and data access belong in the baseline product. Cloud vendors should keep them at least six months without an extra fee. I will not pay to see the evidence. cisa.gov/resources-tools/res…
2
Open source maintainers are expected to secure critical infrastructure with volunteer economics. If a company depends on a project for revenue, contribution should be a cost of doing business.
Fair or impossible?
memorysafety.openssf.org/mem…
5
Vetted researchers can receive internal data from very large platforms when the work studies systemic risk. The status requires a research affiliation and a promise to publish the results free of charge. Access without publication is not scrutiny.
digital-strategy.ec.europa.e…
2
Very large platforms must assess systemic risk every year and take an independent audit on that cadence. Within three months they publish the assessment, the audit, and what they changed. digital-strategy.ec.europa.e…
3
Among people who use AI agents, about 70 percent say a task got faster and 69 percent say they are more productive. Only 17 percent say the team collaborates better. Personal speed is not a team result.
survey.stackoverflow.co/2025…
2
OpenAI rates GPT-6 Cyber Critical for cybersecurity, up from High for GPT-5.6 Cyber. ExploitGym is 48.5 percent versus 32.1. It stayed in bounds on 99.93 percent of runs without the deployment prompt, and 100 percent with it. It is authorized research. openai.com/index/introducing…
1
29
Content moderation transparency should include error rates. A platform that reports takedown volume but hides false positives is publishing activity, not accountability.
Which metric would expose the truth?
digital-strategy.ec.europa.e…
5
