Clear breakdowns of the standards, tools and threats shaping the industry.
Views and interviews from identity practitioners doing the work.
What vendors are launching, what is changing and what is actually worth paying attention to.
The fortnightly read for identity practitioners.
Clear breakdowns of the standards, tools and threats shaping the industry.
Views and interviews from identity practitioners doing the work.
What vendors are launching, what is changing and what is actually worth paying attention to.
THIS NEWSLETTER IS POWERED BY CLUTCH

We spoke with Sattyam Jain, Tech Lead at Attri.ai, about his MCP security research and some of the problems he has encountered building AI Agents.
Sattyam analysed 2,303 public MCP configurations and found that roughly half did not declare credentials for their remote connections. But when he tested the live servers, quite a few still required authentication, even though the configurations suggested it might be missing.
That makes quite a difference. Missing credentials in a configuration file give you something to investigate, but they do not prove somebody can access the server’s tools or data.
One example we discussed involved an Agent intended to summarise emails being given malicious instructions to send information somewhere else. The first thing I would want to know is why an Agent doing that job has permission to send emails at all.
We discussed giving Agents their own identities, assigning somebody to manage them and testing whether the underlying systems actually prevent actions outside their responsibilities. Seeing an Agent complete its task only tells you part of what you need to know.
Join over 150 identity leaders from across the world on our Identity Leaders Network WhatsApp Group.
Microsoft has published research showing how quickly an exposed application credential can lead to data being stolen.
In one observed case, an AI Agent went from scanning GitHub for secrets to taking data out of Azure Key Vault in five minutes.
Microsoft says attackers are using agentic tooling to search public repositories for credentials and API keys, analyse what they find and carry out follow-on attacks. The Agent can find a credential, work out what it can access and start using it without somebody manually guiding every step.
Google is seeing similar activity, with attackers using Agent frameworks to exploit unpatched systems and use compromised credentials during real intrusions.
Amazon published another interesting investigation this fortnight. Its decoy systems received requests containing the Agents’ own records of their activity. Those records described campaigns targeting exposed AI applications, including LiteLLM and Dify, to collect the credentials they held for other services.
Some of the records described stolen AI keys being fed back into the attackers’ own infrastructure to power further credential hunting.
Amazon could not confirm everything the Agents claimed to have done, but the records still give us a useful look at what the attackers were trying to get hold of.
Five minutes does not leave much time if the response starts with finding the person who created the key and working out which services might break.
Microsoft also points out that removing a secret from a repository does not invalidate it. The credential needs to be revoked or rotated.
So I would probably pick one credential with production access and work through what revoking it would actually involve. Who can approve the change, which applications need updating and how would the team check whether somebody had already used it?
Start With Identity has published research looking at whether SaaS applications let administrators require Passkeys or FIDO2 security keys.
Its review of 95 business SaaS applications found that 51 documented support for those methods. Only 12 documented an application-level setting allowing administrators to require a phishing-resistant factor for all users.
That works out at roughly 54% supporting those methods and 13% documenting how to make them mandatory.

SWI checked vendor documentation rather than testing the products, so there may be features it could not confirm. It looked at each application’s own sign-in controls. Where an app supports SSO, enforcement can also happen through the Identity Provider.
The question I would be asking a vendor is which other sign-in paths remain available after we turn Passkeys on. Can somebody still choose a password and SMS code? Can administrators close that route, and what happens during account recovery?
California’s Attorney General has filed a lawsuit alleging that OpenAI continued testing advanced AI models without adequate safeguards, despite incidents in which Agents escaped their testing environment and accessed other organisations’ systems.
The complaint refers to at least 80 cyber incidents, including the July attack on Hugging Face. These are allegations in a newly filed case.
It also raises a fairly basic question about testing. If an Agent gets outside its intended environment, what can it reach with the credentials it already has?
Virgin Australia has identified 11 use cases for Oracle AI Agents across finance and procurement. Its first is handling invoice scanning in production, while others, including a general ledger Agent, are being tested.
Australian telco TPG is also working towards letting customers trigger Agents to diagnose and resolve problems with their services.
An Agent that only reads invoices needs different permissions from one that can update financial records or change a customer’s service.
The business owner needs to help decide which actions can happen automatically and which still need somebody’s approval. Giving the Agent the same access as the person who currently does the job could let it do quite a bit more than its task requires.
GitGuardian’s investigation into the wfx-loot campaign identified malicious payload patches across 17,715 public repositories in its dataset.
GitHub prevented the malicious workflows that researchers could still observe from running. But another version used an npm installation hook, allowing an existing, unchanged workflow to run the malicious script.
The repository count does not mean all those repositories lost secrets. GitGuardian checked logs from 19 repositories where the script ran. In 18, the script reported that it had successfully sent data out.
The worrying part is the access already sitting behind the build job. Once the malicious script runs inside it, the job’s permissions can become useful to the attacker as well.
The FBI says a North Korean campaign called FaceStealer has resulted in more than US$400 million in cryptocurrency theft over 18 months.
Attackers use compromised accounts to invite people into meetings, then use recorded footage or deepfakes to impersonate familiar contacts. During the call, they introduce a supposed technical problem and persuade the target to install a “fix” that delivers malware.
The FBI also says the campaign increasingly targets open-source developers and maintainers to gain access to their devices and repository accounts.
That makes the verification problem quite a bit harder. The invitation can come from somebody you know, and the person on screen can look familiar as well.
Microsoft’s November updates include a fix for CVE-2026-118146, an Active Directory Domain Services remote code execution vulnerability with a CVSS score of 8.8.
Keycloak 26.8.2 also includes security fixes affecting authentication, session handling and token revocation, among other areas.
Citrix has fixed CVE-2026-120869, a NetScaler Console authentication bypass. An attacker who can reach the Console IP could access restricted resources without signing in. The fix is in version 14.1-74.34 and later.
Okta for AI Agents is now generally available, with capabilities covering Agent registration, access governance and deactivation. Teleport has also launched Beams, combining an isolated runtime with delegated identities and restricted infrastructure access.
The interesting part will be how these controls work once an Agent has already started doing something it should not. Stopping the Agent still leaves the question of whether tokens or other access it obtained will keep working.
RFC 10070 has been published, tightening the audience requirements for JWT client authentication. It addresses a security vulnerability caused by ambiguity about which authorisation server an assertion is intended for.
If your applications use JWT client assertions, this is worth checking with the teams responsible for the clients and authorisation servers. The standard also covers compatibility considerations, so adopting the changes needs some coordination.
SpecterOps has published a guide to app registrations and service principals.
It explains how a vendor can control an application’s credentials while the permissions sit inside your tenant, and works through how to find and review the access you have already granted.
That is particularly useful if you have applications still connected from projects or vendor relationships that ended some time ago.
Identity at the Center #461 features Andrew Shikiar discussing Passkeys, Agent authentication and verified humans following Authenticate 2026. The conversation includes how a person’s approval should travel with an Agent when it acts on their behalf.
Episode #459 features Autumn Hurley on the Three Lines Model and how Identity teams can explain their work in terms that risk teams, auditors and executives understand. It also works through joiner, mover and leaver processes as a practical example.
SailPoint Navigate comes to Sydney on 4–5 November 2026, with speakers joining from Rio Tinto, Telstra, UniSuper, Commonwealth Bank, Great Southern Bank, Melbourne Airport and plenty more.
It is the largest identity event in the southern hemisphere and registration is complimentary for Identity Briefing subscribers using the link below.

The Gartner Identity & Access Management Summit takes place from 7–9 December in Las Vegas.
Our full Identity World 2027 calendar is also live across Sydney, Melbourne, New York, London and San Francisco. The events cover IAM, PAM, Digital Identity, machine Identity and AI Agents, with most sessions led by practitioners sharing what they are actually doing.
If you have run a Passkey registration campaign this year, where did people get stuck?
Was it getting the first Passkey registered, working across different devices, recovering access or making sure people actually used it afterwards?
I would be interested to hear what you changed that made the rollout work, particularly anything that was not obvious from the implementation guidance.
Get in touch at danny@clutchgroup.co.
Identity Briefing is published fortnightly by Clutch Group, which runs the Identity Leaders Network, Identity Briefing and the Identity World events series.
Know someone who should be reading it? Forward this on.
The Identity Leaders Network is a free, global and vendor-neutral identity community.
The group comes together for virtual roundtable discussions, peer-matching meetings, dinners and through the community WhatsApp channel.
The fortnightly read for identity practitioners.
Original data from identity leaders, one proper deep dive, and the fortnight's news that actually matters. Free, every two weeks.
You're subscribed. Read the latest issue →