Proof other people can check. Security reviews for AI agents and internet-facing systems. [email protected]witnessops.com Warsaw, PolandJoined September 2025
Swarmchasers (14 Oct): agents on Tencent Cloud made 2,600+ tries at Amap place data. When blocked, some tried API keys left in public code.
A source close to Tencent says it was an evaluation, now ended, with no external breach found.
Which of your keys could an agent find?
Stripe's updated Services Agreement (§1.7): if you use an AI agent to access Stripe, you are "solely responsible for each action initiated by or through the AI Agent," and those actions are "legally binding."
For one payment action: what approved it, and where is that written?
Today (16 Oct) is CISA's KEV due date for Ivanti Connect Secure CVE-2026-104935.
One line in the alert worth keeping: Ivanti's Integrity Checker "may not detect" exploitation, because the commands execute in memory.
If you patched: where is the evidence of what ran before?
MAS says SAFR v1.1 (industry reference, not a new rule) puts a checkpoint before an AI agent’s financial action: identity, mandate match, then a record of the decision.
Useful test: can someone outside the team see that the approval or deny came first?
@ben_callow A vendor PASS is not a control. The useful artefacts are the ones you named: permission boundary, action log, and a stop path — written so someone outside the team can still check them after the dashboard moves on.
Nadella (10 Oct): every meaningful model action should leave "tamper-proof human readable evidence," and an authorised person must be able to stop a model mid-task.
Checkable question: does each consequential agent action leave a record someone else can still read?
@rubengarciaes Agreed. One test worth adding: can someone outside the team rebuild that chain for a single action (which agent, under whose authority, what it could reach) without logging into the vendor dashboard?
@SCMagazine@britive1 Task-scoped, time-bound access also leaves something checkable afterwards: what the agent could reach at the moment it acted, not only what the policy said that week.
@helpnetsecurity@PCISSC Defining which agent actions need approval is the easy half. The half reviewers ask for is a record that each approval came before the action, one someone outside the team can check. Worth noting the PCI guidance is advisory.
Which agent performs the action?
Ars Technica reports prompt injections relayed between MCP-connected agents: a lightly guarded helper passes instructions to an agent that trusts it.
For any consequential action: which identity executes it, and what approval record reaches it?
Evidence survives the dashboard.
A WitnessOps review ends in written findings, the evidence behind each one, and the limits of what was checked. Something a person outside the team can read and check.
@AIsystemsRadar The sandbox answers what it can reach at config time. The harder question after an incident is what it actually did reach, under which identity, and who approved it. Worth recording per action, somewhere the agent can't edit.
@AInDotNet Agreed. And when something does get through, the record matters: which identity ran the action, what it was allowed to do, who approved it. Written so someone outside the team can check it, not just the vendor's dashboard.
116 Followers 243 FollowingBuilding AI agents. Researching AI safety and security. Covering new models, breakthroughs, vulnerabilities, and what actually works.
116 Followers 243 FollowingBuilding AI agents. Researching AI safety and security. Covering new models, breakthroughs, vulnerabilities, and what actually works.
44 Followers 119 FollowingBreaking down how AI policy and regulation shape enterprise decisions. Practical analysis for leaders navigating AI risk, governance, and compliance.
2K Followers 919 FollowingCEO | Al Strategy Executive | Keynote Speaker →
Focused on protecting people from online threats like Al impersonation, deepfakes, fraud, and scams.
3K Followers 4K FollowingAI is full of confident claims. I’m interested in the conditions that make them true. Notes on research, agents, and real-world AI.
2K Followers 1K FollowingNon-developer shipping production AI software. Open-source: The Curator, ØØT. Vice Dean @COTRUGLI. From Lab to Life — field notes, not vendor marketing.
612 Followers 2K FollowingSoftware Architect | 18 yrs in Backend, AI & Data Systems | I explain Architecture, AI Engineering & real product building with simple visuals.
256K Followers 1K FollowingFounder & CEO of @haveibeenpwned, the most trusted name in data breach intelligence. Speaker, blogger, Microsoft Regional Director. Gold Coast, Australia.
36K Followers 35 FollowingNossa, vamos comprar!🫸🫷💥
Os gêmeos dos testes virais
Clique em "seguir" para não cair em cilada
👇Links dos produtos👇
https://t.co/UuD55vGZB6
29K Followers 300 FollowingCreator of WinUtil • GitHub's most popular Windows PowerShell utility | Tech content creator focused on Windows & Linux optimization | https://t.co/MJlDO9ejLc
1.8M Followers 2 FollowingWe're an AI safety and research company that builds reliable, interpretable, and steerable AI systems. Talk to our AI assistant @claudeai on https://t.co/FhDI3KQh0n.
7K Followers 2K FollowingAdvancing AI security by uniting innovators, researchers, builders, and defenders to share insights on securing AI-powered and agentic applications.
469K Followers 2K FollowingInvest in the Unexpected || Founder of @AltCryptoCast, @AltCryptoTalk, @KollectAgency, @48HoursAI & @UseWallyAI || DM for Business Enquiries 💰
225K Followers 3K FollowingFull-time Crypto and Stock trader 🇨🇦| I post charts and everything Crypto and Stocks | Click to join the community Support :[email protected]
55K Followers 89 FollowingKişisel görüşlerimi ve kripto piyasasında yaptığım yatırımları içermektedir. Türkiye'nin en büyük kripto yatırımcılarından biri olarak tanınıyorum.