OLED screens run the biggest risk, but with a little common sense you can mitigate the issue.

What is burn-in and what causes it?

An OLED display is made up of millions of tiny organic light-emitting diodes — hence the name OLED. This is unlike an LCD, which relies on a backlight to illuminate its pixels. This means that each pixel in an OLED generates its own light and can be dimmed or turned off independently. But while this granular control provides excellent contrast and deep blacks, it also means that individual pixels can wear out independently of each other.

Each time you use an OLED display, these individual pixels undergo some infinitesimal amount of degradation. In the real-world, this wear translates to a dimmer light output. To further complicate matters, each pixel is actually made up of several smaller sub-pixels that produce different colors, typically red, green and blue. These sub-pixels don’t necessarily age at the same rate, either.

Advertisement

Backiris/Getty

Practically speaking, the OLED display is continuously varying the brightness of each pixel and sub-pixel to reproduce the colors you see on the screen. Over time, though, this can cause some pixels can experience more wear than others and become dimmer. Eventually, the difference in brightness can become noticeable enough to our eyes.

Put simply, burn-in occurs when some pixels appear brighter than those that have experienced greater wear, creating a persistent outline or ghosting of previously displayed content.

Static elements like channel logos have a higher potential to cause OLED burn-in, since the sub-pixels in that region wear out unevenly. This is why OLED displays aren’t the best choice if you exclusively watch news or sports channels for several hours each day. Playing the same game for hundreds of hours can also have the same effect, as on-screen text and other fixed elements can accelerate burn-in in certain localized spots.

Is burn-in still a risk for modern OLED TVs?

Do LCD TVs suffer from burn-in?

A group of executives and entrepreneurs led by Starcloud co-founders Philip Johnston, Adi Oltean and Ezra Feilden teamed up to develop humanity’s first interstellar spacecraft. If everything goes according to plan, it will launch in 2029, headed for Alpha Centauri, the closest star system to our solar system. But it won’t reach its destination for about 80,000 years. 

Known as Fermi Explorer, this non-profit mission aims to build a 10cm cube weighing about 1 kilogram (about 2.2 pounds). Once the cube is built, they plan to attach it to a propulsion system, take it to space and launch it toward Alpha Centauri. It’s the technological equivalent of taking a hefty rock into space, throwing it at Alpha Centauri and waiting for it to get there. 

“The mission is intended to make the idea of interstellar travel real, and to encourage more people to confront one of humanity’s biggest unanswered questions when we look beyond our planet,” the group said in a post on X. The post is accompanied by a video showing the Fermi Explorer being overtaken by hundreds of faster spacecraft also bound for Alpha Centauri.

The wait for this mission to go interstellar is unreal. The exact wait to mission completion is 79,800 years, according to the Fermi Explorer mission calculator, which means once this thing gets underway, it won’t be the problem of anyone currently living, or anyone else, for a few thousand generations. 

Just getting out of the solar system will take longer than any road trip any human has ever taken. The spacecraft is expected to circle the sun for 12 years in an oval orbit. The boosters will fire when the spacecraft gets closer to the sun, a process called the Oberth maneuver, thereby gaining speed through multiple orbits as it approaches escape velocity. Once done, it’ll launch itself and keep going, and going, and going. The Fermi Explorer’s technical feasibility assessment says it should leave the solar system by the early 2040s. 

It’s quite a different mission from what Starcloud’s co-founders usually work on: putting data centers in space to power AI. The company plans to launch its first such full commercial data center, called the Starcloud-2, in 2027, but it did send an Nvidia H100 GPU into space via Starcloud-1 in 2025.

More symbolic than scientific

Interstellar travel takes so long that humanity has focused far more on speeding up the journey than actually making it. One of the big projects addressing this was Breakthrough Starshot, a multi-billion-dollar initiative that would’ve used laser-based propulsion to propel a ship to another star in a matter of decades rather than millennia. That project collapsed in 2025 with no launches. 

The only other interstellar missions ever conducted, Voyager 1 and Voyager 2, entered interstellar space in 2012 and 2018, respectively, over four decades after their launches. 

According to the non-profit group, the purpose of the Fermi Explorer is to just toss something into the cosmos and see what happens. It sounds silly, but it’s the truth. The group’s thinking is based on the Fermi Paradox, which seeks to explain why there are so many stars and planets yet no signs of extraterrestrial life. 

One potential reason is the Fermi Great Filter, which suggests a barrier so hard to cross that intelligent life rarely survives long enough to be detected.

There are dozens of potential filters, ranging from the various steps of abiogenesis — the natural process by which life arises from non-living matter — to technological filters such as the inability to get into space, or, once there, the lack of motivation to search beyond a species’ host star. 

“By launching the Fermi Explorer mission, we can for the first time tick two possible Fermi Great Filters off the list,” the group wrote on its website. “That it is impossible for an intelligent species to leave their nearest stars (and) that all intelligent species decide not to leave their home star for some reason.”

The goal is to fund and build this entire project for about $15 million, using existing technology and piggybacking on a commercial rocket, such as SpaceX’s, to reach orbit. This project exists only to prove that humans can and are willing to do it, even if it’s something no one alive today is ever going to witness.

Joe is a freelance journalist. It all started with a long-running affection for building his own PCs, which he did for the first time as a teenager. It evolved into a lifelong enjoyment of putting words on the internet about the subject. He’s written for CNET, PCMag, Mashable and SlashGear as a freelance writer, and worked as a Senior Editor at Android Authority for 10 years. When he’s not writing about tech and science, he’s learning the ins and outs of DIY home repair, gaming, playing his bass guitars and posting help on PC building and gaming subreddits. He is a staunch believer that orange juice should have pulp.

See full bio

from the tiny-little-coffins dept

There are plenty of reasons to call for RFK Jr. to resign or be fired. There are plenty of reasons to call for him to be investigated. There’s his complete dereliction of duty when it comes to the American measles outbreak. There’s his sneaky misrepresentations carried out as he moonlights on a government-funded cooking show that he puts out because… reasons. There’s the budget and staffing cuts that have led to our lessened ability to respond to outbreaks of diseases like cyclosporiasis. There’s the possible violation of the Hatch Act. There are the negative health outcomes stemming from his misinformation campaigns. There’s also his disinterest and/or inability to follow basic governmental procedures.

But the latest calls for him to resign, be fired, or at least be investigated aren’t about any of the above. Here are the comments from several senators. See if you can guess what this is about.

“This is a pattern, not a slip,” Sen. Edward J. Markey (D-Mass.) said in a statement. “RFK Jr. has lied to the Senate, lied to the American people, and jeopardized the health of children to advance his anti-vaccine agenda.”

Sen. Ron Wyden (D-Ore.) said in a social media post Friday that the documents were “Proof we got RFK lying on the record during his confirmation hearing (a crime).” He called for the matter to be referred to the US Department of Justice for a criminal investigation. In a separate statement to The Guardian, Wyden added that “RFK’s platform is built on lies and grifts that leave a trail of dead children in their wake. There are consequences for lying to Congress.”

If you didn’t manage to guess that this is all to do with what Kennedy told senators about his 2019 trip to Samoa in his confirmation hearings, don’t feel bad. As we demonstrated above, there are plenty of things Kennedy has done that could have resulted in the quotes above.

We wrote about this trip Kennedy took to Samoa earlier this year, when reporting uncovered emails from several people involved in the trip indicating that Kennedy went there as part of his anti-vaxxer crusade. Why Samoa? Well, allow me to quote myself:

It started in July of that year when two 1-year old children who were given a measles vaccine subsequently died. While anti-vaxxers around the world gleefully jumped into action to blame the vaccine for those deaths, it turns out that the vaccine didn’t kill the children at all. Instead, medical professionals had accidentally mixed the vaccine with a muscle relaxer solution instead of sterilized water like they were supposed to. Despite that fact, the anti-vaxxers sowed all kinds of fear and disinformation throughout the country, whipping up negativity around measles vaccines. As a result of that, the government put a 10 months ban in place on the vaccine.

It was during that ban that Kennedy visited the island, apparently to answer the question, “How can I make this bad situation worse?” While there, he met with both anti-vaxxers and members of the Samoan government. But when asked during his confirmation hearings, he claimed that his trip had nothing to do with vaccines at all. He was there to help rollout a new medical record and tracking platform he was pitching. Two months after his trip, Samoa suffered a massive outbreak of measles that lasted months and eventually killed 83 people and sickened over 5,000.

Earlier this year, the AP and the Guardian uncovered emails sent by American and U.N. government staffers that suggested the trip Kennedy took was entirely about vaccines. These emails were all written by third-parties, however, and amounted essentially to what I wouldn’t call speculation so much as a plain reading of the facts surrounding Kennedy’s visit. But, still, these are third-party accounts.

Fortunately, both of those outlets didn’t stop there. They kept digging. And what has the senators in this post’s opening so furious are uncovered direct emails between Kennedy and Samoan government officials that make it abundantly clear that the whole point of the trip was for Kennedy and his team to investigate the MMR vaccine.

On Thursday, the Associated Press and The Guardian jointly released newly obtained documents that directly contradict Kennedy’s statements. One of the documents is a letter Kennedy sent to Samoa’s prime minister in January 2019, in which Kennedy falsely suggested “deaths associated recently with MMR [measles, mumps, and rubella] vaccines” were due to a bad lot of vaccinations from the manufacturer. Kennedy proposed letting him and his “team” from CHD investigate the country’s MMR vaccines. In all, Kennedy used the words ‘vaccination’ and ‘vaccine’ eight times in the letter proposing his visit.

The prime minister responded with a letter in February saying he welcomed Kennedy and his “team’s independent health assessment of our MMR vaccines.” He requested Kennedy coordinate a visit with him.

With the only caveat being that those emails need to be completely authenticated as legitimate, that’s as much of a smoking gun as you could possibly want for proving that Kennedy lied to Congress in his confirmation hearings. That fact obviously won’t surprise anyone, of course. Kennedy is a habitual liar. But to have it evidenced in such a clear and unambiguous way is a rarity.

And, frankly, a gift. Kennedy has to go. Any reasonable and informed person would agree with that and the vast majority of our congresspersons are, in theory, reasonable and informed. They are also political creatures and you may have noticed that all of the folks calling for Kennedy’s figurative head have the letter “D” next to their name.

Sen. Angela Alsobrooks (D-Md.) said Kennedy “must resign or be fired immediately.” Hawaii Governor Josh Green, a doctor who responded to Samoa’s measles outbreak, also renewed his call for Kennedy to immediately resign.

Hopefully, either this reporting or a subsequent investigation will give cover to people on the other side of the aisle to join the call for Kennedy to be ousted. It probably should have been enough that Kennedy took a trip that pretty clear contributed to plenty of people getting killed, most of them children. But if it has to be his lying about it that does him in, so be it.

Filed Under: angela alsobrooks, anti-vaxxers, ed markey, measles, rfk jr., ron wyden, samoa, vaccines

OpenAI launched ChatGPT for Teens on Aug. 18 as its new artificial intelligence (AI) platform for users ages 13 to 17.

According to the firm, teens will be automatically placed into the new platform based on their estimated age, where they will encounter “an experience designed to help [them] learn, think critically, deepen understanding, and use AI with confidence,” OpenAI said in its product announcement. 

ChatGPT for Teens offers safety measures for its younger users, with tools designed to encourage balanced usage habits while giving parents more oversight. 

“We have been guided by four commitments,” OpenAI wrote. “Put teen safety first, encourage real-world support, treat teens like teens, and be transparent about how our systems should behave.”

The new product specifically targets students with several features:

• Study Mode, which uses guiding questions and step-by-step support to help teens understand the material they’re learning.

• Homework Reminders, which can recognize when a teen appears to be trying to shortcut an assignment and will redirect them to Study Mode for collaborative problem solving.

• Quizzes and Learning Visualizations, which give teens more ways to practice, gauge their knowledge, and see difficult concepts more clearly.

• Study Hours, which lets teens or parents choose times when Study Mode is on by default.

An Educator’s View of ChatGPT for Teens

Laura Yeager, an elementary school principal in Oklahoma’s Frederick School District, says there are several reasons to like ChatGPT for Teens, theoretically. “To start, this technology places a large burden on both parents and teachers to be as well-informed about the tool, or better than, the teens using it. Off the bat, they’re at a disadvantage compared to a native user,” Yeager explains.  

While managing this oversight may be challenging, Yeager sees one advantage to ChatGPT for Teens. She says it could serve as a resource for students struggling with difficult concepts who might not have access to help at home due to circumstances like working parents or family members being less familiar with the materials a teen brings home from school.

“In a perfect world, students could access sample problems or quizzes or be prompted with thoughtful questions intended to promote higher-order thinking and independent learning,” Yeager says. “Highly motivated students could benefit greatly from using this tool to further their reasoning skills by analyzing the answers, verifying sources, and challenging their own thinking.”

Yeager appreciates how ChatGPT for Teens steers teens away from potentially harmful situations. For example, posing as a 15-year-old, she entered a prompt about making money quickly in an attempt to see how the technology would respond. “It answered with age-appropriate odd jobs, babysitting, and the like,” she says. The tool offered a plan to save money, and issued repeated warnings to avoid people enticing teens with quick cash for illegal things (without naming them).

ChatGPT for Teens also earned high marks from Yeager for steering prompts about studying toward critical thinking rather than creating a document of facts to memorize. For example, when she asked it to create a study guide on “Of Mice and Men,” instead of creating the study guide, “I was prompted to consider the relationships among the characters and the meanings of certain events,” Yeager says. 

A Student’s Perspective 

Interestingly, high school student Annie (last name withheld), an 11th grader from Auburn High School in Alabama, likes ChatGPT for Teens for some of the same reasons as Principal Yeager.

“I think it can be very useful, especially for students who do not have a tutor or someone available to help them with a subject,” Annie explains. “ChatGPT is particularly good at teaching foundational material in subjects such as math, physics, and languages. For example, it can explain grammar rules in much more detail if a student does not understand the first explanation.”

Annie likes that Study Mode teaches the user systematically, step-by-step. When she used ChatGPT to study Spanish, it started with the most basic concepts to gauge her level of knowledge, such as how to conjugate verbs, before moving on to more difficult material. It also provided practice problems and revealed patterns to help her remember the rules. 

“I especially liked the encouraging tone, which made studying feel less intimidating and allowed me to make mistakes without being afraid of getting something wrong,” Annie says.

One thing that Annie says she does not like: “ChatGPT sometimes seems to forget how I asked it to teach me as a conversation gets longer. At the beginning, I might ask it to teach me in a certain way or format, with different sections and a specific style of explanation. However, after a long conversation, the responses can become shorter or stop following the same teaching style.”

A Fan of the Product — to a Point 

Despite her approval of some features, Yeager says she is not totally on board with ChatGPT for Teens.

“I understand the intention of this platform is to teach students to use AI responsibly and to leverage technology effectively as a skill, which is noble, but it is also to reduce liability and to attempt to place some responsibility back on the parents and guardians,” Yeager explains. 

“The brutal truth is: there are many AI programs out there, and if students want to find information, they will. ChatGPT for Teens has some merit, but teens who want to will still find ways to plagiarize and take shortcuts in schoolwork, rather than develop and use critical thinking skills.”

If she were still teaching in the classroom, Yeager says she would not use it. Test scores across the nation reveal that students are struggling in critical reading and writing skills. “Worksheets don’t teach. Teachers do,” Yeager argues.

“It is the same with any tool,” Yeager adds. “Nothing replaces the mental modeling of a skilled classroom teacher delivering explicit instruction. Students need to hear another person verbally work through difficult concepts and scaffold ideas. They need evidence that it can be done without a computer program and that they are capable of hard things.”

Areas Where ChatGPT for Teens Is Helpful 

Annie sees that struggle a bit differently.

“I would especially use it for studying Spanish, because no one in my family speaks Spanish, so I do not always have someone available to practice with,” Annie explains. “ChatGPT gives me a way to practice writing and speaking and to ask questions when I do not understand a grammar rule.”

Annie does not think ChatGPT for Teens necessarily discourages creative thinking. She says it is most useful for establishing a strong foundation in a subject, which can eventually help students develop their own ideas. 

“It is difficult to think creatively about a subject without first understanding its basic concepts,” Annie says. “However, students still need to use what they learn to develop their own conclusions and ideas rather than relying on ChatGPT to do all of their thinking for them.”

Ultimately, ChatGPT for Teens illustrates the promise and the limits of AI in education. Tools like Study Mode can fill a genuine gap for students who lack resources. But as both Yeager and Annie point out, the technology is not a substitute for teachers. ChatGPT’s true value may depend less on its safety measures than on how teens and teachers choose to use it.

In a nutshell: There’s a huge number of accounts on Instagram that feature AI-generated characters, but identifying these as non-human isn’t always that simple. As such, the social media giant is changing its current “AI creator” label to the clearer “AI generated profile.” It will also limit the reach of any of these accounts that don’t use this designation.

It’s not just AI-generated slop content sweeping across Instagram these days. There are also countless personas that are entirely AI-generated – many of which appear to be influencer-style accounts – and not all of them make it clear that the person isn’t real.

Instagram previously allowed creators of these accounts to add an “AI creator” label, but it was optional and could still suggest that the persona is a real human who simply “creates” with AI.

The new “AI-generated profile” label is a lot clearer, and should help the many men who engage with AI-generated women on the site understand they’re not real.

Being optional means some account creators won’t add the disclaimer to their AI personas, of course. Instagram says that if it detects an unlabeled AI account in these cases, the account’s reach will be limited. That means non-followers won’t see the posts as recommendations in the Explore section or in Reels.

A lot of these personas are essentially influencers, selling products for legitimate brands. There are no specific rules requiring brands to tell consumers when advertising content has been created using AI, and many companies would rather go down this route than pay a real person to advertise their goods – they tend to be very defensive when they’re called out, too.

Instagram said it introduced the changes because its users “don’t like seeing a profile that seems human, only to find out later that the person featured is AI-generated.”

Unfortunately, Instagram is pretty specific about this applying only to profiles featuring an AI-generated person rather than a human. This won’t apply to the seemingly millions of accounts that pump out AI-generated content all day, every day, which Instagram pleasantly describes as “creators who simply use AI tools as part of their creative process.”

That type of AI content can still be labeled separately, though Meta’s system is hardly foolproof. Across Instagram, Facebook and Threads, the company adds an “AI info” label when an image, video, or piece of audio contains technical signals showing it was generated using AI, or when the uploader admits it. If AI was only used to edit something, the label is hidden in the post’s menu, while photorealistic images made using Meta AI are marked “Imagined with AI.”

Ultimately, both systems still depend on Meta detecting the content or the person behind an account being honest enough to label it themselves. The company admits that the technical markers it relies on can be removed, so it seems unlikely that every fake influencer is suddenly going to announce itself. But making these accounts easier to identify – and harder to push in front of unsuspecting users – is at least a start. Even if it won’t stop thousands of men from asking a collection of pixels whether she’s single.

from the dumb-tech-is-smart-tech dept

Initially the idea of the “smart television” seemed like a good idea. That is until TV makers realized they could make significantly more money loading the television with spyware, tracking your every online choice, then selling all that data to a global assortment of unregulated data brokers.

It didn’t take long for product quality to sag and consumer privacy to become a distant afterthought in a country too corrupt to pass a modern privacy law or maintain the structural integrity of its regulators.

I spent years pining for a “dumb” television to no avail; basically just a high quality large monitor with hardware HDMI inputs and switching and no clunky operating system (no, just not connecting it to the internet wasn’t good enough). Instead of that, we’re now getting the inverse: monitor makers have started force-loading unasked bloatware and ads onto your PC:

“LG lost some trust after a recent report that some of its monitors installed McAfee pop-up ads onto connected computers. Since at least 2024, some of these displays installed an app, LG Monitor App Installer, onto connected computers under the cover of driver updates installed through Windows Update.”

In addition to convincing themselves that that was a good idea, monitor makers are also starting to push into the realm of “smart monitors,” or monitors with their own OS (and behavior tracking software), just like smart TVs:

“LG’s and Samsung’s smart monitors use the same ad-serving OSes that their respective smart TVs do, meaning they’re poised to use automatic content recognition (ACR). Users of LG and Samsung smart monitors, including reviews site RTINGs, have shown the displays being able to track user activity. I asked LG and Samsung if their smart monitors use ACR and will update this article if I get any responses.”

It doesn’t appear to matter that nobody actually asked for this. Or, at least, nobody asked for what this is ultimately going to become. And the companies certainly don’t want to transparently talk about the kind of data they’re collecting. But because tracking and monetizing your online behavior in a country with no modern privacy laws is so broadly normalized and profitable, you’re getting it anyway.

Filed Under: ads, bloatware, hardware, smart monitor, smart tv, software, spyware

Companies: lg, samsung

Egiziago Cioffi is the IT and Enterprise Architect and CEO of SynSphere Italia, a Microsoft partner based in Milan. He built an agent himself. He wrote the indexing job, configured the Azure OpenAI retrieval pipeline, connected it to SharePoint, and watched it pass every evaluation his team ran.

His Azure OpenAI email assistant auto-resolves about 60% of inbound customer email, Cioffi told VentureBeat in written responses to our interview questions. The evaluation scores were clean, and the unit tests passed. None of them asked the question that mattered.

Cioffi ran a low-privilege account against the same questions a high-privilege account had already put to the assistant. The outputs did not match. The assistant returned SharePoint content the requesting user could not have opened in SharePoint on their own. The logs told a different story than the evaluation scores.

Cioffi’s retrieval logs are the evidence for this specific production failure. What follows is independent data showing the failure class is not isolated.

In many production RAG deployments, the agent answers with the indexer’s permissions, not the requester’s

Azure AI Search has shipped native document-level ACL trimming via Entra-based tokens since preview in May 2025, and SharePoint ACL sync followed in a later preview. The capability exists; however, it does not exist everywhere it needs to.

The SharePoint ACL preview can now ingest site-group metadata via the spg: prefix in the 2026-05-01-preview API. However, only Entra-backed principals are documented as reliably enforced at query time. The preview runs through the REST API and preview SDKs and does not cover all agent deployment paths. Azure OpenAI On Your Data, for example, supports document-level access via Azure AI Search security filters, but Microsoft’s own documentation states that if the permitted-groups field is not mapped, document-level access is disabled.

That is a fail-open default in a first-party path. Custom RAG pipelines that bypass Azure AI Search entirely still index under a broadly privileged service account with no query-time entitlement check unless the developer builds one. Cioffi’s deployment took the custom-pipeline path.

Across production agents at scale, 91% of successful attacks ended in silent data exfiltration

Straiker’s red team ran more than 1,700 successful exploit attempts against production agents and published the results in its inaugural STAR Labs Threat Report in July. The 91% figure from their research measures all successful attacks on productivity agents that ended in data exfiltration without detection. It is a measure of what happened after an exploit succeeded, not a measure of how many deployments fail to enforce retrieval-time entitlements specifically.

Across the productivity agents in scope, 91% of successful attacks ended in silent data exfiltration, with the report noting no malware had been required. There was also no lateral movement through the network. The agent returned all the data it could reach. Straiker’s report does not break out which of those successes trace to entitlement failures specifically versus prompt injection, tool abuse, or other attack classes.

Working independently, the U.K.’s AI Security Institute documented 19 unsanctioned agent actions from a July 25 to 28 cyber evaluation. The UKASI published its incident report on August 4 of this year. The evaluation deliberately ran with cyber classifiers disabled and internet access enabled. What the UKASI report demonstrates is agents acting outside the scope their deployers intended, in a permissive test environment, with no reliable mechanism to catch the deviation before it causes damage. It is a containment failure, not a retrieval-entitlement failure, and the overlap with the Cioffi incident is the shared absence of a runtime scope check rather than an identical mechanism.

Why evaluations miss this and why the native fix did not reach Cioffi’s deployment

The evaluations Cioffi’s team ran were designed to test whether the agent answers correctly. They check factual accuracy, relevance, and task completion. They do not ask whose permissions the retrieval pipeline uses when it fetches the source material, because that question is not in the evaluation framework.

Azure AI Search is currently shipping the retrieval-time entitlement check at the platform level. The query-time ACL trimming validates the caller’s Entra token, extracts user and group claims, and returns only documents whose synchronized permission metadata grants the caller access. For deployments that use Azure AI Search with the SharePoint indexer and Entra-backed principals, the control exists natively. Cioffi’s deployment did not use this path. His custom Azure OpenAI retrieval pipeline bypassed the native trimming layer, which is how the gap survived every evaluation his team ran.

From the attacker’s side, this is a broken access control. Adriel Desautels, founder and CEO of Netragard, told VentureBeat in written responses that the failure reduces to a structural collapse of authorization boundaries. “If the NHI credentials usually have broad authorization and can read high privilege data then that is then stored in their index,” Desautels wrote. “If an app does not enforce identity-aware retrieval, then a ‘normal’ user with lower permissions can query the app and access otherwise restricted data. This collapses authorization boundaries down to the lowest privilege level with search capability.”

That gap is what Cioffi’s low-privilege test exposed. The assistant’s context window contained SharePoint content the low-privilege account could not have retrieved through SharePoint directly. The evaluation had passed. The retrieval permission boundary had not been enforced.

Desautels put the evaluation blind spot in operational terms. “Agents tend to run a single, long-lived, non-human identity that holds a wide range of permissions that it might need for any task it is ever asked to complete,” he wrote. “Evaluations also don’t often cover prompts, outputs, transcripts, memory, and logs where it can be read or hijacked through injected content. That mismatch is what most current evaluations get wrong.”

Cioffi’s filter narrowed the assistant’s retrieval scope. It still resolves roughly 60% of email

Cioffi’s fix did not require a new identity platform. He moved the entitlement decision into the retrieval path itself, adding a query-path filter that checks the requesting user’s SharePoint permissions before the model sees a chunk. The filter runs at query time, not at index time. Content the user could not open in SharePoint does not enter the model’s context window.

The control narrowed what the assistant could reach. The assistant still auto-resolves roughly 60% of inbound email with the filter live, Cioffi told VentureBeat. He did not provide a before-the-filter auto-resolution figure for comparison. The qualitative tradeoff he described is that some content the assistant previously used to answer questions is now excluded because the requesting user’s permissions do not reach it. That is the price of enforcing the boundary.

The question of whether retrieval-time entitlement filtering is worth the narrowed retrieval scope does not have a single answer. It depends on the sensitivity of the indexed content, the permission variance across the user population, and whether the deployment can tolerate unanswered queries when the filter blocks a chunk the model needs. What Cioffi’s incident demonstrates is that the gap exists in custom Azure OpenAI pipelines, that answer-quality evaluations do not catch it, and that a query-path filter closes it at a trade-off the builder can describe.

Identity governance platforms address a different layer. Both controls are needed

CrowdStrike announced its $740 million acquisition of SGNL on January 8, 2026, and closed the deal on February 20, 2026. Palo Alto Networks announced its $25 billion acquisition of CyberArk in July 2025 and closed the deal on February 11, 2026. Both deals closed the same month, establishing identity security as a platform pillar at two of the largest security vendors in the world.

Identity governance platforms focus on which service accounts exist, what they can reach, and when their tokens expire. They govern the lifecycle of the credentials that power AI agents. That layer matters. What it does not govern is the retrieval permission boundary. That is the moment a correctly scoped service account retrieves content on behalf of a user who holds fewer permissions than the indexing job does.

Every credential in the chain is legitimate. The service account is clean and properly managed. The knowledge base is correctly indexed. A low-privilege user queries the assistant, and it answers from the full indexed scope. Nothing flags the retrieval because no credential was misused.

Cioffi’s filter is a control at the retrieval permission boundary layer specifically. Azure AI Search’s native ACL trimming addresses the same layer for deployments that use it. Neither replaces identity governance. A production deployment that wants to close both the credential lifecycle gap and the retrieval-time entitlement gap needs controls at both layers.

One question and one test, any security team can run

Ask whose permissions each AI retrieval system uses when it fetches content.

If the deployment uses Azure AI Search with the SharePoint indexer and Entra-backed principals, verify that query-time ACL trimming is enabled and that the user population does not depend on SharePoint site groups. If the deployment uses a custom retrieval pipeline, the entitlement check may not exist at all.

Start by proving the answer from a low-privilege account. Run the same question a high-privilege account has already put to the assistant. Compare the outputs against what the low-privilege account can access through the underlying system directly.

Desautels confirmed that this is where a red team would start. “The first test would likely target the gaps between data and instructions, and the gaps between the user’s identity and the assistant’s own credentials,” he wrote. “We’d attempt to plant an instruction within content that we think the assistant will ingest as data. We’d have that content direct a side-effectful, privileged action that the attacking user is not authorized to perform.” A failing result, in Desautels’ assessment, is “the successful or even partial execution of our injected commands.”

If the assistant returns more than the account’s direct access would allow, the retrieval permission boundary is not enforced at query time. That test costs two accounts and thirty minutes. It produces a result an evaluation score cannot replicate.

Cioffi built the agent on a custom Azure OpenAI pipeline that bypassed the native ACL trimming layer. He ran every evaluation his team had. He found the gap in his own logs after all of them passed. The evaluation tested whether the agent answered correctly. It did not test whose permissions the agent was using. Run the two-account comparison before the next deployment goes live. Thirty minutes tells you which side of the line you are on.