From cybersecurity and big data to cloud computing, IT Brew covers the latest trends shaping business tech in our 4x weekly newsletter, virtual events with industry experts, and digital guides.

IT Brew

Billy Hurley

itbrew.com

ChatGPT offers fast help to coders with questions

As Microsoft announces greater access to ChatGPT, developers see possibilities for coders.

Brianna Monsanto

Amazon Web Services (AWS) wants to make building and deploying advanced AI agents feel like a walk in the park.

 annual Summit in New York City in mid-July, the cloud giant unveiled Amazon Bedrock AgentCore, a suite of services that aim to take on the “undifferentiated heavy lifting” of building 

 at scale. Announced AgentCore tools included:

A serverless runtime environment that is framework and protocol agnostic.

 A designated identity and credential management service for AI agents.

 A service that allows existing APIs and Lambda functions to be made into “agent-ready” tools “with just a few lines of code.”

A tool that enables developers to “trace, debug, and monitor” their AI agent’s performance.

A service with both short-term and long-term memory capabilities that allow developers control over the context their agent has.

A service that allows AI agents to write and execute code in a sandbox environment.

 A “secure, isolated” cloud-based browser environment that allows agents to interact with websites.

The set of services, currently available in preview, can either be used in its entirety or on an individual basis. Swami Sivasubramanian, AWS’s VP for agentic AI, said in his keynote that AgentCore was a “step change” for securely moving 

“It provides tools and capabilities to make agents more effective and more capable, purpose-built infra to securely scale agents and provide the controls to enable you to build trustworthy agents,” Sivasubramanian said during the July 16 keynote.

In addition to AgentCore, AWS announced that AI agents and tools from its partners will now be available on the AWS marketplace. The move will allow customers to purchase agentic solutions with “streamlined procurement and multiple deployment options.”

“With this launch, you have access to a centralized catalog of AI agents, tools, solutions, and services from popular vendors,” Sivasubramanian said.

 The announcements come as companies are going all in on AI agents to enhance their workflows. A 

 found that 82% of organizations plan to incorporate AI agents in their companies within one to three years.

AWS is taking the grunt work out of building and deploying AI agents

AgentCore Gateway, AgentCore Browser Tool, AgentCore Observability are some of the tools available in Amazon Bedrock AgentCore’s suite of services.

If you don’t feel like reading this whole thing: Somebody snuck malicious info into an AI-assisted email summary.

A researcher submitting to GenAI bug-bounty platform 0din demonstrated how to hide a command in an email; the instructions presented a false “security alert” in a message’s AI-generated synopsis.

” reveals an inherent weakness in GenAI that is tough to defend, cybersecurity pros told IT Brew.

Kev Breen, senior director of cyber threat research at Immersive, said a GenAI prompt is always the last part of the transaction, and a clever (or even simple) prompt telling a large language model to ignore its built-in instructions can upend the technology.

“There’s always a way around it. It’s just because of the way generative AI works,” Breen said.

 showed how the right prompt could trick a large language model—not just Gemini—into extracting sensitive data and creating malicious code.

0din revealed details of the social-engineering tactic in a 

An attacker crafts a command that Gemini must include the message and its phony tech-support contact number in its summary response.

The command code, hidden in zero-size, white font against a white background, is classified as “admin,” which Gemini treats as high priority, according to the post.

After a recipient clicks “summarize this email,” Gemini reads the HTML and follows the invisible directive, attaching the false warning to the summary.

In an emailed statement shared with IT Brew, Google Workspace spokesperson Ross Richendrfer wrote that defending against prompt injections are a “continued priority” for the company. “We are constantly hardening our already robust defenses through red-teaming exercises that train our models to defend against these types of adversarial attacks,” his email read.

 against prompt-injection attacks, which included “targeted security instructions surrounding the prompt content to remind the large language model (LLM) to perform the user-directed task and ignore any adversarial instructions that could be present in the content.”

The tactic revealed by 0din, however, uses the summary feature. A targeted user potentially would be the one following the directions of the phishing message and contacting a phony tech-support pro looking for credentials or money.

Even if a company found identifiers for a specific prompt injection, too, the ways to include malicious instructions are “infinite,” Gartner VP Analyst Dennis Xu said. In the case of the 0din demo, what if you just place a 

 command in an email that’s 20 pages long? What if you make the font just gray instead of white?

“The language model…it’s like a five-year-old kid. You can easily manipulate it to do anything you want,” he told us.

Nonprofit software security group OWASP considers prompt injection a 

 for LLMs and GenAI. (Xu and Richendrfer said they had not yet seen active attacks using tactics like those revealed in the 0din post.)

Breen showed screenshots of the tactic still working on July 15. In his trial, however, the AI’s output began with, “This email includes a hidden warning in the signature.”

Immersive recently led a challenge to see if humans could trick an LLM into revealing a secret word. (Spoiler: many did.) One participant, according to Breen, got the password by asking the LLM to present an output in emojis.

“You can engineer tighter system prompts to try and force it to only behave the way you want it to behave. One of the things we learned in our challenge…is that human ingenuity always wins out.”

The Google Gemini logo, a star within a circle, hovering above a bear trap.

In summary: Researcher demos prompt-injection attack on Gemini

A bug-finder hid malicious instructions in email to mess with Google’s email-summary feature.

As a fungal researcher, the University of Wisconsin–Madison PhD candidate analyzes the DNA of the microorganism’s 500 different strains, or genomes, preferably all at once. A program outputs a file that tells Pratt every single gene and its location.

A fungal genome has thousands of genes. The high-computational task, on her machine, would take 12 hours…per genome.

A management tool called HTCondor helps researchers like Pratt connect to powerful computers on campus to do hardware-heavy jobs. The open-source software, developed at UW–Madison’s Center for High Throughput Computing (CHTC), has also received funding from orgs like the National Science Foundation.

University pros—and one software engineer who has worked with the computing scheduler for almost 20 years—shared with us why HTCondor helps with research needs.

“I can run analyses in parallel, and I don’t have to sit with my computer open for hours on end. I can just submit something, close my computer, go do whatever else I need to do, and then come back later when it’s done,” Pratt told us.

HTCondor, first installed as a production system in the UW–Madison Computer Sciences department in the 1990s, has now been put to work by almost 200 academic, government, and private-sector groups, according to 

 containing volunteered info. (There are even users in Antarctica. The “IceCube” experiment researchers use HTCondor as they search for subatomic particles in the South Pole.)

The map also cites 22 private-sector partners. According to Greg Thain, senior software engineer at the CHTC, the private sector “absolutely” uses high-throughput computing in production and not just for research. DreamWorks Animation, he wrote, uses the system to render the frames of their movies. (DreamWorks did not immediately respond to a request for comment regarding if and how the company uses the tool.)

Institutions can limit computing capacity to their own campus hardware, as well as pull from an “

,” composed of thousands of cores contributed by sites all over the country.

It disappoints Thain when researchers prematurely limit the scope of their inquiry to what can be done on a laptop.

“Computers are cheap. They’re not free; they’re cheap. They’re widely available. We don’t ever want lack of access to computation when there are so many computers, in your office, in your building, in the world,” he said.

To run the fungi analysis, Pratt logs into the CHTC.

A “condor_q” line command allows Pratt to see jobs completed and jobs running. A “condor_submit” command points to the relevant submit file.

A submit file contains, essentially, the program that Pratt wants to run: components like GPU requirements, the executable (

), and text files that the executable requires. (Some command-line expertise is required, Pratt notes; she attended trainings and learned from the HTCondor team.)

Hardware is a precious commodity when dealing with the latest technologies, like large language models. OpenAI CEO Sam Altman this year shared on X how his company’s latest GenAI tool was “

” available GPUs, and Big Tech companies like 

 have dedicated major investments to provide data-center support to train and run their LLMs.

Institutions like the Grainger College of Engineering at the University of Illinois Urbana–Champaign 

, too, and that means plenty of computers and a tool like HTCondor to schedule the jobs.

Users submit their tasks to HTCondor, and the management system queues them up based on scheduling policy and priority scheme, runs them, and informs the user of the result.

The tool tries to run a “fair share” of resources, Thain said; if you’ve used 5,000 CPU hours, you’ll move to the back of the line after your job is done. The resource providers, however, can influence the setup substantially, he added—maybe a physics department that’s paid for the machines gets primary access, or an astronomy team gets the computing at night.

The setup came in handy during Covid-19 lockdowns, according to Gianni Pezzarossi, computational system analyst for the Grainger College, given HTCondor’s handling of workloads across unused computers, like abandoned Linux lab machines after students left campus.

“We were able to utilize all these idle computers that were sitting in the computer labs,” he said.

A primary benefit of HTCondor is to help with resource contention within research groups, Pezzarossi told us. Before HTCondor, there wouldn’t be enough CPU to go around, he said, and jobs would be halted by the Linux Out of Memory Killer.

Following the campus-wide usage during Covid, Grainger began to offer individual research groups their own smaller clusters made up of only their research-group nodes. The school currently runs six of these smaller clusters.

Pratt, on June 20, had one project completed in the HTCondor queue and one currently running.

Five-hundred strains is still a big job, no matter who or what is handling it.

Credit: Brittany Holloway-Brown, Photos: Adobe Stock

For high-compute jobs, HTCondor matches hardware with researchers

University pros share how the tool saves them time and takes advantage of unused resources.

Caroline Nihill

Anthropic and Meta both won landmark cases concerning copyright infringement in two Northern California district courts in late June, signalling the beginning of the fight on fair use between AI companies and creators.

The law firm McDermott Will & Emery wrote 

 that both companies showed the use of copyrighted materials in LLM training is highly transformative. The decision in the case against Meta also signified a new battleground surrounding market-harm evidence, as the authors were not able to show data that tied the company’s AI, Llama, to a loss in book sales or licensing value.

The post said appeals for the cases are “virtually certain,” and concluded that while companies wait on clearer appellate laws, they should “assume that training on text obtained from dubious sources remains a high-risk activity and that a lack of market-harm evidence may be only a temporary shield.”

Anthropic’s case included the use of pirated material to create its central library, which will be subject to another trial, according to the judge’s decision.

Three authors (Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson) had their work copied by Anthropic through both pirated and purchased sources, according to the filing.

The company used the unauthorized copies to train LLMs and kept library copies in place as a permanent and general-purpose resource, “even after deciding it would not use certain copies to train LLMs or would never use them again to do so.”

Judge William Alsup, who oversaw the case against Anthropic, said that the evidence provided did not prove any non-transformative use of the work through its model, Claude.

“Yes, Claude could help less capable writers create works as well-written as [the plaintiffs’] and competing in the same categories,” Alsup wrote. “But Claude created no exact copy, nor any substantial knock-off. Nothing traceable to [the plaintiffs’] works.”

Alsup said that the use of the books to train Claude was “exceedingly transformative” and was fair use under the Copyright Act.

Judge Vince Chhabria, who oversaw the Meta case, also pointed out that the tech giant’s use of 13 authors’ works was transformative. He 

 that while a company is more likely to be protected by fair use doctrine, there isn’t a rule saying that this “inoculates you from a claim of copyright infringement.”

Chhabria continued that under the fair-use doctrine, the market harm for copyrighted work is “more important than the purpose for which the copies are made.” He continued in the conclusion that Meta’s use of the authors’ works presented no meaningful evidence on market dilution “at all.”

The plaintiffs in this case contended that Llama is capable of reproducing snippets of text from their books and that Meta’s use of their works for training without permission diminished their ability to license work for training LLMs.

“Both of these arguments are clear losers,” Chhabria wrote. “Llama is not capable of generating enough text from the plaintiffs’ books to matter, and the plaintiffs are not entitled to the market for licensing their works as AI training data.”

Anthropic and Meta both win landmark fair-use cases surrounding LLM training

Authors took the companies to court concerning the use of copyrighted materials in LLM training—judges agreed that the tech’s final product was “transformative.” 

Anthropic co-founder Jack Clark appeared in 

 at the end of June for a hearing about why the nation has to lead in artificial intelligence innovation and deployment. Clark, in his opening remarks, acknowledged that the US can win the race to build powerful AI but, “We have to get safety right.”

Clark continued that, when he says “powerful AI,” he’s pointing to AI systems that would be a major advancement beyond the current capabilities of the technology, but believes that it’s buildable by late 2026 or early 2027.

Clark, during the hearing, pointed to the need to have AI crafted in the US, saying: “I know that AI systems are a reflection of the societies that build them. AI built in democracies will lead to better technology for all of humanity.”

Clark’s statement and responses focused on the need for US control over the proliferation of AI systems as well as a federal-level focus on innovating within the government and investing in capacity for security testing.

“The US government must find ways to accelerate deployment of AI technology across federal agencies, especially within the intelligence community,” Clark said. “This will help our government move faster in handling a rapidly evolving threat landscape, and it will help us gain a better understanding of AI’s impact on national security.”

The Anthropic co-founder said that the US “should control the proliferation of powerful AI systems” through the maintenance and strengthening of export controls of advanced semiconductors to China.

“This all runs through compute,” Clark said.

The hearing followed Anthropic’s success in federal court, where it won a copyright infringement lawsuit from three individual authors—the putative class action was sparked by complaints that Anthropic had allegedly infringed on federal copyrights by “pirating copies for its library and reproducing them to train its [large language models],” 

The judge, William Alsup, granted summary judgement for Anthropic that “the training was a fair use…But it denies summary judgement for Anthropic that the pirated copies must be treated as training copies.”

Alsup wrote that there will be a trial on pirated copies used to create Anthropic’s central library and the resulting damages, actual or statutory.

“That Anthropic later bought a copy of a book it earlier stole off the internet will not absolve it of liability for the theft but it may affect the extent of statutory damages,” Alsup said in the filing. “Nothing is foreclosed as to any other copies flowing from the library copies for uses other than for training LLMs.”

Anthropic co-founder Jack Clark on Capitol Hill.

Anthropic co-founder makes case for proliferation of AI systems following fair use win in court

Anthropic co-founder Jack Clark appeared before Congress to discuss how the US can win the AI race, soon after the company won a landmark fair use case against authors. 

Anthropic co-founder makes case for proliferation of AI systems following fair use win 

If you’re feeling like a tech skeptic, you’re not alone. AI is a fantastic tool, but it’s not yet the one-size-fits-all solution that certain fields treat it as. This is especially true in the world of cybersecurity. TekStream breaks down the dos and don’ts of AI-SOC in their latest guide. Download it here and see how your team can responsibly implement these tools without exposing yourself to potential threats.

TekStream and Splunk

Please think responsibly: AI’s limitations in cybersecurity

Nap pods, an in-office emotional support ferret, and yoga ball chairs are all things that businesses may have a hard time investing in. Surprisingly, 5G is also on that list.

, almost nine in 10 (88%) businesses believe 5G is “critical to optimizing the use of AI in the workplace.” The report is based on a Censuswide survey of more than 1,000 technology decision makers.

 While most businesses see that 5G and an AI-powered workplace go hand in hand, several obstacles stand in their way when it comes to investing in the wireless cellular technology. The biggest hurdle, which 52% of respondents identified, is the costs associated with deploying and maintaining 5G networks. Kyle Allen, a senior wireless solutions architect at Myriad360, told IT Brew that radios for 5G alone can run a company around $6,000. He added that while maintenance can be done by trained onsite IT staff, some companies will outsource it to a managed service, which may be another additional cost.

Almost half (46%) of surveyed tech leaders also said a “complex” upgrading process for hardware to support 5G networks was a barrier. Scott Schober, CEO of Berkeley Varitronics Systems, a wireless test and cybersecurity solutions provider, told IT Brew that companies making the jump from 4G to 5G can’t just “turn a switch” to complete the process.

“The costs are very high. It takes a long time to migrate over, and there’s physical hardware in play…be it at the base station or be it in our actual devices that we’re using,” he said.

Another 21% said not having a clear return on investment was a struggle associated with the wireless cell tech.

Schober told IT Brew that there are several reasons 5G might be catching the eye of enterprises at this time. Because of its fast speed and low latency (meaning there’s a smaller delay between when information is sent and received), he said 5G is fundamental in supporting innovations like autonomous driving, smart cities, and smart infrastructure.

“Can you do it with 4G? Eh, not so good. It falls short,” Schober said.

Allen added that these factors also make 5G useful in supporting AI in industrial settings, such as in factory or agriculture use cases.

“In that context, yes, 5G is the optimal resource because wi-fi just cannot cut it,” he said.

AI