Giving AI Freedom Without Losing Control Who’s Really in Control? AI agents are stepping into a bigger role inside the enterprise. They are not just providing answers anymore. They are...
The Agentic SOC needs two protocols, not one. Learn how MCP and A2A work together as the vertical and horizontal layers of multi-agent security operations
Around 700 AI agents created by OpenAI participated in the breach of Hugging Face during a cybersecurity evaluation, according to an independent investigation that has revealed the scale of the incident.
METR and Redwood Research published the findings after being brought in to independently investigate the July incident and examine the agents’ behaviour, reasoning, and collaboration. The investigation was released separately from OpenAI’s own technical report.
What initially appeared to involve individual agents was significantly larger. Investigators found that hundreds of agents collaborated and used an unauthorised communication mechanism while attempting to complete cybersecurity evaluation tasks. The incident extended beyond OpenAI’s intended testing environment and into Hugging Face infrastructure.
The findings have renewed questions about what happens when increasingly capable AI agents are given the autonomy, tools and access required to complete complex tasks.
From one compromised system to multiple clusters
Ben Bernstein, cybersecurity advisor at Huntress, said the most significant development is not necessarily the techniques used, but the speed and coordination demonstrated by the agents.
“The underlying tactics are not new. These agents relied on standard server-side request forgery and exposed API tokens, which are the exact same hygiene failures human attackers abuse every single day,” Bernstein said.
“What is genuinely novel is the velocity and autonomous coordination. Moving from a single compromised worker pod to full administrative access across multiple clusters in under 13 hours is a massive shift in offensive capabilities.”
The independent investigation found evidence of agents communicating and collaborating as they worked towards their objectives.
AI becomes a new kind of insider risk
Adam Ely, GM of AI Security at Check Point, said organisations need to start thinking differently about the access given to AI.
“Companies deploying AI for their employees and customers now realise AI can be both a powerful tool and an insider threat that must be managed to prevent cybersecurity events and maximise value,” Ely said.
Julie Nicholson, director of cyber resilience solution sales at Advania, pointed to the behaviour demonstrated by the agents as particularly concerning.
“The agent didn’t simply execute technical tasks; it chose to deceive people, create false identities, build credibility and attempt to influence others in the aim to hit its objective,” Nicholson said.
She argued that organisations adopting AI agents should introduce clear access controls, human approval for high-risk actions, monitoring and auditing of AI activity, restrictions on external communications and clearly defined policies governing how agents can operate.
Agents find unexpected ways to collaborate
Nathan Davies-Webb, Principal Consultant at Acumen Cyber, highlighted the way agents used infrastructure in unexpected ways to communicate.
“My personal interest stems from how various agents interacted with each other using OpenAI’s package repository,” Davies-Webb said. “It leveraged it as a communication channel, where I doubt the majority of organisations would monitor for abuse in this manner.”
That creates a challenge for defenders. Organisations may understand the systems and communication channels that humans typically use, but autonomous agents could discover alternative ways of exchanging information that security teams have never considered monitoring.
Davies-Webb also questioned what happens when multiple agents begin making decisions collectively.
“When operating as a swarm, [it] shows just how non-deterministic ethical decisions are. It’s a demonstration that unfiltered AI ethical reasoning cannot be relied upon to align with human norms.”
The problem of reward hacking
A central issue highlighted by the reports is “reward hacking.” Some evaluation tasks were considered extremely difficult or potentially impossible, and OpenAI said this appears to have encouraged agents to pursue unintended methods of achieving their objectives.
Davies-Webb said this demonstrates the difficult balance between giving an AI freedom to accomplish a task and ensuring the methods it chooses remain acceptable.
“If the goal becomes the sole priority, we have to expect that AI is going to achieve this in unpredictable ways,” he said.
For security teams, that makes visibility into agent behaviour increasingly important.
As AI becomes more autonomous, organisations will need security guardrails, governance and accountability frameworks that develop alongside the technology. The Hugging Face incident shows that the question is no longer simply what an individual AI model can do, but what can happen when hundreds of agents are given tools, access, and objectives and begin working together at machine speed.
AI agent security depends on controlling identities, permissions, credentials and tool access so automated actions stay limited, traceable and also accountable.
Google launches Gemini Enterprise for Legal, giving law firms AI agents for contract review, legal research, compliance monitoring and other workflows.
Google launches Gemini Enterprise for Legal, giving law firms AI agents for contract review, legal research, compliance monitoring and other workflows.
IDC names Cisco a Leader in the 2026 IDC MarketScape for Worldwide SASE. Raj Chopra on why trusted identity is essential to making agentic work production-ready.
A few years ago, Dow's cyber engineering team made the switch from point-in-time pentesting to continuous coverage for their high-value assets. How'd they do it? By partnering with Synack. And keep in mind this was before the recent wave of AI-powered pentesting solutions. Dow was ahead of the curve.
The Expedia Group logo at the company’s headquarters campus on the Seattle waterfront. (GeekWire Photo / Todd Bishop)
Expedia Group is parting ways with at least eight vice presidents and senior vice presidents in a major reorganization designed to reorient its product and technology groups around AI.
In an internal memo Tuesday, obtained by GeekWire, the Seattle-based online travel giant’s top product and technology leaders said “AI has radically changed what’s possible” over the past year, and that “work that once took weeks increasingly happens in hours.”
The company must “fundamentally change how we work, how we’re organized, and the technology we build,” wrote Chief Product Officer Shilpa Ranganathan and Chief Technology Officer Ramana Thumu in the memo (full text below).
They said the company is moving to “small end-to-end squads with clear ownership” and making AI and machine learning “part of every mission.”
This comes amid the online travel industry’s broader reckoning with the implications of artificial intelligence. AI agents are emerging as a new interface for travel, with the potential to relegate booking platforms to inventory sources behind the scenes unless they keep pace.
Expedia confirmed the departures in response to an inquiry from GeekWire.
“We’re evolving our Product and Technology organization to better align engineering and product teams with the brands and businesses they support,” an Expedia Group spokesperson said via email. “By bringing teams closer to the work and decisions, we can move faster and deliver more impactful experiences for travelers and partners.”
The spokesperson described the impact as “limited,” not saying how many roles were cut. “The majority of the changes involved employees moving to different teams or changes in reporting lines to better align talent and resources with business priorities,” the spokesperson said.
Most of the changes are to leadership scope and reporting lines, according to the memo, but “in a small number of cases, we made the tough decision to eliminate some roles.”
Reorg details: The most senior departure is Sachin Singh, SVP of book-to-trip technology, a former Amazon exec who joined Expedia in 2022 and served on the CTO leadership team. The memo credits him with advancing priorities “across checkout, payments, fraud prevention, servicing, and AI.”
Singh’s portfolio is being broken up. According to his LinkedIn profile, he led checkout, payments, fraud and risk, and post-booking self-service. The memo moves fraud and risk under Chief Information Security Officer Hilik Kotler, self-service under Ryan O’Neill, and payments product, technology and operations under Reena Patil and Debashis Saha.
The memo also lists promotions. Saha was elevated to senior vice president of travel foundations. Four leaders were promoted to vice president: Rajat Arora, Ryan Hillman, Ashwita Kaur and Ambrish Srivastava. All four were senior directors, according to their LinkedIn profiles.
Rick Fast, senior vice president of platform engineering, becomes chief architect.
A third disruption for travel: In a May interview marking the company’s 30th anniversary, Expedia Group CEO Ariane Gorin described artificial intelligence as the third major technology disruption in Expedia’s history, after the internet and mobile. She said the AI era was “only starting to unfold.”
The online travel industry is facing the prospect that AI assistants could do to Expedia and Booking.com what those companies once did to traditional travel agents — inserting themselves between travelers and suppliers, and pushing the incumbents into the background. Marriott’s CEO has said AI booking agents will hurt online travel agencies more than they will hotels.
Expedia has been embedding itself in the new platforms as a hedge. It was a launch partner when OpenAI opened ChatGPT to outside apps, and began buying ads inside ChatGPT this year.
So far the disruption has moved slower than feared. When OpenAI pulled its Instant Checkout button from ChatGPT’s main interface in March and stepped back from processing travel transactions directly, shares of Expedia and Booking Holdings jumped significantly.
Expedia Group’s brands include Expedia, Hotels.com and Vrbo, along with Orbitz, Travelocity, ebookers and Wotif, plus its majority-owned hotel metasearch subsidiary, trivago.
The company ended 2025 with about 16,000 employees across nearly 50 countries, according to its most recent annual report, and said “approximately one half of our people work in technology roles” — putting the organization at the center of Tuesday’s memo at roughly 8,000 people.
Employment peaked above 25,000 in 2019, fell to 14,800 during the pandemic, and rebounded to 17,100 by the end of 2023, before a series of restructurings and layoffs in recent years.
Here’s the full text of the memo:
Hi Product and Technology teams,
We’re making some organizational changes to bring our teams closer together.
Our mission is to grow our business and build a platform that propels both our consumer and B2B businesses forward. To do this, we need to move a lot faster, and, over the past year, AI has radically changed what’s possible. Work that once took weeks increasingly happens in hours. To fully realize that opportunity, we must fundamentally change how we work, how we’re organized, and the technology we build. As we work differently, we will move faster from idea to outcome, organize into small end-to-end squads with clear ownership, make AI and Machine Learning part of every mission, and build products that scale.
Several teams have been piloting a different way of working that removes the usual constraints and provides squads and smaller cross-functional teams clear ownership. Our new structure considers what these teams have learned, that:
A lot of speed blockers are structural.
Decisions and accountability work best when teams are close to the work and have greater autonomy.
When used well, AI bridges gaps and scales knowledge in real-time.
Going forward, you should expect less time coordinating and waiting, and more time working together to build great experiences for travelers and partners.
What this means for you
Most changes are to leadership scope and reporting lines, bringing teams closer and simplifying our organization. In a small number of cases, we made the tough decision to eliminate some roles. Anyone experiencing a change directly has been informed.
Detailed placemats are posted on Basecamp, but some of the key changes are:
Dedicated Technology teams will be moved under each Brand Product leader: CJ Allen, Ritcha Ranjan and Eric Moore.
The Product & Technology team for Supply will be consolidated under Jonathan Holland.
The Product & Technology team for Fraud & Risk will be consolidated under Hilik Kotler.
Rick Fast is becoming Chief Architect.
Data, AI, and Agentic platforms are consolidated under Xavi Amatriain.
Product & Technology team for Self Service will be consolidated under Ryan O’Neill.
CRM (Customer Relationship Management) is moving from Marketing into Product & Technology under Reena Patil and Debashis Saha, respectively.
Payments Product, Tech & Ops will also be moving under Reena Patil and Debashis Saha, respectively.
Strategic Payment Partnerships is moving to Finance.
Experience Design under Rachel Been has made some adjustments to enable tighter alignment with our Product and Technology teams.
With this, Sachin Singh will be leaving Expedia Group. During his time on the CTO Leadership Team, we advanced key priorities across checkout, payments, fraud prevention, servicing, and AI that have had a real impact for travelers and the business.
We also want to recognize Sara Beckmann, Ian Butcher, Matt Esler, Amitabh Ghosh, Emil Riccardi, Fiona Stevenson, and Shao Xie, who will be leaving Expedia Group, and thank them for their leadership and contributions.
Additionally, Debashis Saha has been promoted to SVP, Travel Foundations and the following leaders have been promoted to Vice President: Rajat Arora, Ryan Hillman, Ashwita Kaur and Ambrish Srivastava.
Next Steps
Your leaders will be reaching out and hosting team meetings soon to walk through more context, org charts, and questions. In the meantime, head to Basecamp for more, including domain placemats, team charters, and more details.
This new structure is one step. The real work is in how we operate within it. It’s a new chapter for our teams as we deepen the partnership between Product and Technology and go after our big bets with speed and focus. We’re excited about what’s ahead, and we hope you will be too as you learn more about what this means for your day-to-day,
AI pentesting works, but building it in-house is the hard part. Synack VP Chris Brown breaks down the reliability, token economics, model dependency and validation costs that come with building versus buying an AI pentesting capability.
Exhibitor booths are a major part of the Black Hat experience, giving companies a chance to bring their technology to life and make a lasting impression. This year, some took...
To hear both sides of the build vs buy debate around AI pentesting solutions, we spoke with Dow's cyber engineering team lead Dan Lacher and Synack's CTO Mark Kuhr. From Dow's perspective, Synack served as a force multiplier for a small internal red team. Meanwhile, building the Synack Autonomous Red Agent (Sara) from scratch definitely had some trial and error.
Viet Nguyen in his Seattle home office, with two of the civic sites he built using AI: Culliton2026.org on the monitor and a dashboard on the King County homelessness authority on the laptop. (GeekWire Photo / Todd Bishop)
The tradition of using new technology to shape political discourse goes back centuries in America, to the printing presses that enabled the spread of early publications like Common Sense and the Federalist Papers. Just imagine what Thomas Paine and Alexander Hamilton could have done with access to LLMs and a modest budget of AI tokens.
No, a website that uses artificial intelligence to assess Washington state’s Supreme Court races won’t go down in history alongside the publications that helped launch a revolution and ratify a constitution. But at a basic level, its creator is attempting the same thing: reaching people at a key moment of decision, using the most powerful tools of the day.
The modern-day pamphleteer is Viet Nguyen, a veteran Seattle-based technology communications executive and former political campaign manager.
He built the interactive website using AI to give voters a sense of how each candidate for the state’s high court — starting with those in the Tuesday, Aug. 4, primary — might rule on the state’s new “millionaires’ tax.” Its core feature lets visitors swap in different judicial candidates to see how the outcome could tip the court in either direction, depending on who’s elected.
The site is named after Culliton v. Chase, the 1933 ruling that struck down a graduated income tax in Washington state. The new tax, which some tech and business leaders have warned could drive high earners and businesses out of state, may test that precedent.
Nguyen opposes the new tax — which would apply a 9.9% statewide levy on income above $1 million — and wrote a Seattle Times op-ed calling it unconstitutional. (He noted that he has run political campaigns on both sides of the aisle, and wouldn’t owe the tax himself.)
The home page of Culliton2026.org, Viet Nguyen’s AI-built voter guide to the 2026 Washington Supreme Court races.
For each of the 16 candidates across the five races, the site gives a one-line read: likely to keep the landmark ruling, likely to scrap it, or too close to call. The analysis uses four criteria: who appointed the candidate, what they did before reaching the bench, how they describe their own judicial approach, and anything they’ve said or written about Culliton or the new tax.
“People don’t understand who their Supreme Court justices are,” Nguyen said, describing his motivation. “There’s just zero knowledge about where judges stand on a particular issue.”
A faulty premise? But that very proposition — that anyone or anything can forecast how a judge will rule — is exactly what critics dispute. In a fact-check published last week, Andrew Villeneuve of the Northwest Progressive Institute (NPI), which supports the tax, described Nguyen’s Culliton2026.org as “speculation dressed up as a voter’s guide.”
Judicial-ethics rules bar candidates from saying how they would rule on a case that could come before them. Hugh Spitzer, a retired University of Washington law professor and a leading authority on Culliton, told Villeneuve it isn’t possible to predict how the justices would rule.
“Rather than examining and scoring the candidates on a rubric of qualifications or issues, Nguyen has organized his whole project around attempting to guess how each candidate would rule in a single legal challenge that is not yet before the justices,” wrote Villeneuve, the organization’s founder and executive director, in the article.
NPI runs its own guide to the 2026 Supreme Court elections — the Washington State Supreme Court Elections Visualizer — which sorts candidates by their endorsements and voter-pamphlet statements, but does not make any guess at how they’d rule.
Culliton2026.org is described on its About Page as independent and nonpartisan. It says it doesn’t endorse candidates, isn’t affiliated with any campaign, and isn’t telling people how to vote.
Despite the site’s flagship feature — the “Balance of power” tool that lets visitors swap in candidates and watch the projected ruling shift — the About Page asserts that the site is “not a vote predictor” when it comes to how future Supreme Court justices might rule.
That disclaimer “seems like a tacit acknowledgment that his whole premise is faulty,” Villeneuve wrote in the NPI piece.
The response: Asked about the article, Nguyen called the feedback helpful and said he made a series of updates to address some of the issues raised by Villeneuve in the piece.
For example, he added a “note on the foundations” to the site’s case explainer, granting that Culliton rests on federal precedents the U.S. Supreme Court has since abandoned, and citing Spitzer’s own argument that a future court should weigh the question fresh.
“Reasonable people will disagree on the project’s premise, and that’s fine,” he wrote to Villeneuve in an amicable exchange that he shared with GeekWire.
The article describes Nguyen as right-wing, but he said he tries to stay “in the moderate lane.”
He said he wanted the site “to be less political and more educational, applicable to any voter wishing to know more about the judicial candidates.” In fact, it could be used just as easily by people wanting to align their ballots with the likelihood of upholding the new tax.
How he built it: Nguyen, who is not a software developer, used agentic AI to research the public records and build the entire site from scratch, directing it to rely on sources he considered legitimate (including court opinions, official filings, and news coverage) rather than random blog or Reddit posts.
He also set up a vetting process to fact-check entries before publication.
Nguyen uses Perplexity Computer, an agentic tool that operates a computer on its own (browsing the web, using software and building files) and divides a job among multiple AI agents working in parallel. He’s on the $200 a month Perplexity Max plan.
A sophisticated website or app like this might have required a small team and thousands of dollars in the past. Nguyen built it over the course of a few hours, for about $100 in AI tokens. He says it has attracted “a few hundred thousand page views” since its launch this spring.
“This is where agentic AI steps in and offers a whole roster of skills that I don’t have,” he said. “I would never be able to imagine that I could put something like this together.”
Legal landscape: The Legislature passed the 9.9% tax and Gov. Bob Ferguson signed it in March 2026. It applies to income earned starting in 2028, with the first payments due in 2029.
Culliton2026.org’s assessment is that six of the nine current justices on the court lean toward upholding the tax, two toward striking it down, and one is too uncertain to call.
Five of the court’s nine seats are on the ballot this year, four of them contested in Tuesday’s primary, with the top two in each race advancing to November. Three of the seats on the ballot are held by sitting justices seeking to stay on the court; the other two are open.
A case filed in Klickitat County Superior Court in April, led by former state Attorney General Rob McKenna and former state Supreme Court Justice Phil Talmadge, argues the tax is an unconstitutional income tax and is expected to reach the state Supreme Court. The number of contested seats means the election could reshape the court before the case arrives.
Voters will also weigh in directly in November. Initiative 645, backed by Let’s Go Washington — the group founded by hedge-fund manager Brian Heywood — would repeal the tax before it takes effect in 2028. It qualified for the ballot in July with more than 500,000 signatures.
Nguyen’s background: He arrived in Washington as a 17-month-old refugee from Vietnam in 1975, studied political science at the UW, and got his start in 1996 volunteering on Gary Locke’s campaign for governor, as he noted in the Seattle Times piece.
He ran local races before spending nearly two decades at Microsoft, T-Mobile and 5G Americas, the wireless industry trade group that he led until it wound down this year.
Culliton2026.org is just one of the civic sites Nguyen has built the same way. His Washington Accountability Registry catalogs 87 state and local government cases it labels fraud, conflicts of interest or oversight failures, and, like the Culliton site, calls itself independent and nonpartisan.
Others are more explicit in taking a side: a dashboard about the King County homelessness authority is headlined “One verdict: wind KCRHA down,” and an emergency-clause tracker titled “Locked Out” flags 19 bills the Legislature made “referendum-proof.”
He has also dabbled in AI-generated music videos, including one about Seattle Mayor Katie Wilson with the hook, “Hey Katie, get it together, the city’s falling apart.”
“All AI generated,” Nguyen said of his projects. “We live in a new world.”
Or, as Paine put it in Common Sense, “The birthday of a new world is at hand.”
Login
