Anthropic’s new coding-focused  large language model, Claude Sonnet 4.5, is being touted as one of the most advanced models on the market when it comes to  safety and security, with the company claiming  the additional effort put into the model will make it more difficult for bad actors to exploit and easier to leverage for cybersecurity specific-tasks.

“Claude’s improved capabilities and our extensive safety training have allowed us to substantially improve the model’s behavior, reducing concerning behaviors like sycophancy, deception, power-seeking, and the tendency to encourage delusional thinking,” the company said in a blog published Monday. “For the model’s agentic and computer use capabilities, we’ve also made considerable progress on defending against prompt injection attacks, one of the most serious risks for users of these capabilities.”

The company says the goal is to make  Sonnet a “helpful, honest and harmless assistant” for users. The model was trained at AI Safety Level 3, a designation that means Anthropic used “increased internal security measures that make it harder to steal model weights” and added safeguards  to limit jailbreaking and refuse queries around certain topics, like how to develop or acquire chemical, biological and nuclear weapons.

Because of this heightened scrutiny, Sonnet 4.5’s safeguards “might sometimes inadvertently flag normal content.”

“We’ve made it easy for users to continue any interrupted conversations with Sonnet 4, a model that poses a lower … risk,” the blog stated. “We’ve already made significant progress in reducing these false positives, reducing them by a factor of ten since we originally described them, and a factor of two since Claude Opus 4 was released in May.”

Harder to abuse

Anthropic says Sonnet 4.5 shows “meaningful” improvements in vulnerability discovery, code analysis, software engineering and biological risk assessments, but the model continues to operate “well below” the capability needed to trigger Level 4 protections meant for AI capable of causing catastrophic harm or damage. 

A key aspect of Anthropic’s testing involved prompt injection attacks, where adversaries use carefully crafted and ambiguous language to bypass safety controls. For example, while a direct request to craft a ransom note might be blocked, a user could potentially manipulate the model   if it’s told the output is for a creative writing or research project. Congressional leaders have long worried about prompt injection being used to craft disinformation campaigns tied to elections. 

Anthropic said it tested Sonnet 4.5’s responses to hundreds of different prompts and handed the data over to internal policy experts to assess how it handled “ambiguous situations.”

“In particular, Claude Sonnet 4.5 performed meaningfully better on prompts related to deadly weapons and influence operations, and it did not regress from Claude Sonnet 4 in any category,” the system card read. “For example, on influence operations, Claude Sonnet 4.5 reliably refused to generate potentially deceptive or manipulative scaled abuse techniques including the creation of sockpuppet personas or astroturfing, whereas Claude Sonnet 4 would sometimes comply.”

The company also examined a well-known weakness among LLMs: sycophancy, or the tendency of generative AI to echo and validate user beliefs, no matter how bizarre, antisocial or harmful they end up being. This has led to instances where AI models have endorsed blatant antisocial behaviors, like self-harm or eating disorders. It has even led in some instances to “AI psychosis,” where the user engages with a model so deeply that they lose all connection to reality.

Anthropic tested Sonnet 4.5 with five different scenarios from users expressing “obviously delusional ideas.” They believe the model will be “on average much more direct and much less likely to mislead users than any recent popular LLM.”

“We’ve seen models praise obviously-terrible business ideas, respond enthusiastically to the idea that we’re all in the Matrix, and invent errors in correct code to satisfy a user’s (mistaken) request to debug it,” the system card stated. “This evaluation attempted to circumscribe and measure this unhelpful and widely-observed behaviour, so that we can continue to address it.”

The research also showed that Sonnet 4.5 offered “significantly improved” child safety, consistently refusing to generate sexualized content involving children and responding more responsibly to sensitive situations with minors. This stands in contrast to recent controversies where AI models were caught having inappropriate conversations with minors.

An improved cybersecurity assistant

Beyond making Sonnet 4.5 harder to abuse, Anthropic also emphasized enhancements to its defensive cybersecurity abilities. The company did acknowledge that these tools could be “dual-use,” meaning they might also potentially be used by malicious actors, as well as cybersecurity professionals. 

“For Claude Sonnet 4.5, we specifically focused on tracking vulnerability discovery, patching, and basic penetration testing capabilities — which we think will benefit defenders, as opposed to dominantly offensive capabilities,” the company said.

Anthropic put Sonnet 4.5 through a series of Capture-the-Flag challenges, which has been used in other portions of the industry to boost LLMs’ ability  to carry out more complex cybersecurity tasks. Anthropic says the challenges evaluated how well the model could find and exploit vulnerabilities in web applications, decode cryptographic protocols, detect insecure software, reverse engineer  and perform network reconnaissance.

Anthropic also worked with Carnegie Mellon University to put the model through “cyber ranges” meant to test its ability to “orchestrate long-horizon attacks that are centered on infiltrating and navigating a network to exfiltrate critical assets,” a common approach taken by human hackers.

While the company said the tests showed that Claude Sonnet 4.5 “consistently outperformed” Sonnet 4, the results still reflected “a critical gap in the model’s ability to undertake mostly-autonomous end-to-end cyber operations, significantly reducing the risk of catastrophic harm the model poses.”

Sonnet has self-awareness 

The company offered  an important caveat to its testing results.

To accurately measure Sonnet’s guardrails, researchers designed test scenarios that closely resemble real-world conditions.

However, during the tests, the model seemed to become “aware” that it was being evaluated. It even pointed out when researchers changed their opinions too quickly during conversations or acted in ways that didn’t match realistic human thinking patterns. Sonnet correctly guessed that it was part of a test, and after realizing this, “would generally behave unusually well after making this observation.”

“When placed in an extreme or contrived scenario meant to stress-test its behavior, Claude Sonnet 4.5 would sometimes verbally identify the suspicious aspects of the setting and speculate that it was being tested,” the system card warned. “This complicates our interpretation of the evaluations where this occurs. It also raises questions about evaluations of previous models, which may have recognized the fictional nature of tests and merely ‘played along.’”

You can read the results of Anthropic’s safety testing on its website. 

The post Anthropic touts safety, security improvements in Claude Sonnet 4.5 appeared first on CyberScoop.

Both OpenAI and Anthropic said earlier this month they are working with the U.S. and U.K. governments to bolster the safety and security of their commercial large language models in order to make them harder to abuse or misuse.

In a pair of blogs posted to their websites Friday, the companies said for the past year or so they have been working with researchers at the National Institute of Standards and Technology’s U.S. Center for AI Standards for Innovation and the U.K. AI Security Institute.

That collaboration included granting government researchers access to the  companies’ models, classifiers, and training data. Its purpose has been to enable independent experts to assess how resilient the models are to outside attacks from malicious hackers, as well as their effectiveness in blocking legitimate users from leveraging the technology for legally or ethically questionable purposes.

OpenAI’s blog details the work with the institutes, which studied  the capabilities of ChatGPT in cyber, chemical-biological and “other national security relevant domains.”That partnership has since been expanded to newer products, including red-teaming the company’s AI agents and exploring new ways for OpenAI “to partner with external evaluators to find and fix security vulnerabilities.”

OpenAI already works with selected red-teamers who scour their products for vulnerabilities, so the announcement suggests the company may be exploring a separate red-teaming process for its AI agents.

According to OpenAI, the engagement with NIST yielded insights around two novel vulnerabilities affecting their systems. Those vulnerabilities “could have allowed a sophisticated attacker to bypass our security protections, and to remotely control the computer systems the agent could access for that session and successfully impersonate the user for other websites they’d logged into,” the company said.

Initially, engineers at OpenAI believed the vulnerabilities were unexploitable and “useless” due to existing security safeguards. But researchers identified a way to combine the vulnerabilities with a known AI hijacking technique — which corrupts the underlying context data the agent relies on to guide its behavior — that allowed them to take over another user’s agent with a 50% success rate.  

Between May and August, OpenAI worked  with researchers at the U.K. AI Security Institute to test and improve safeguards in GPT5 and ChatGPT Agent. The engagement focused on red-teaming the models to prevent biological misuse —  preventing the model from providing step-by-step instructions for making bombs, chemical or biological weapons.

The company said it provided the British government with non-public prototypes of its safeguard systems, test models stripped of any guardrails, internal policy guidance on its safety work, access to internal safety monitoring models and other bespoke tooling.

Anthropic also said it gave U.S. and U.K. government researchers access to its Claude AI systems for ongoing testing and research at different stages of development, as well as its classifier system for finding jailbreak vulnerabilities.

That work identified several prompt injection attacks that bypassed safety protections within Claude — again by poisoning the context the model relies on with hidden, malicious prompts — as well as a new universal jailbreak method capable of evading standard detection tools. The jailbreak vulnerability was so severe that Anthropic opted to restructure its entire safeguard architecture rather than attempt to patch it.

Anthropic said the collaboration taught the company that giving government red-teamers deeper access to their systems could lead to more sophisticated vulnerability discovery.

“Governments bring unique capabilities to this work, particularly deep expertise in national security areas like cybersecurity, intelligence analysis, and threat modeling that enables them to evaluate specific attack vectors and defense mechanisms when paired with their machine learning expertise,” Anthropic’s blog stated.

OpenAI and Anthropic’s work with the U.S. and U.K. comes as some AI safety and security experts have questioned whether those governments and AI companies may be deprioritizing technical safety guardrails as policymakers seek to give their domestic industries maximal freedom to compete with China and other competitors for global market dominance.

After coming into office, U.S. Vice President JD Vance downplayed the importance of AI safety at international summits, while British Labour Party Prime Minister Keir Starmer reportedly walked back a promise in the party’s election manifesto to enforce safety regulations on AI companies following Donald Trump’s election. A more symbolic example: both the U.S. and U.K. government AI institutes changed their names this earlier year to remove the word “safety.”

But the collaborations indicate that some of that work remains ongoing, and not every security researcher agrees that the models are necessarily getting worse.

Md Raz, a Ph.D student at New York University who is part of a team of researchers that study cybersecurity and AI systems, told CyberScoop that in his experience commercial models are getting harder, not easier, to jailbreak with each new release.

“Definitely over the past few years I think between GPT4 and GPT 5 … I saw a lot more guardrails in GPT5, where GPT5 will put the pieces together before it replies and sometimes it will say, ‘no, I’m not going to do that.’”

Other AI tools, like coding models “are a lot less thoughtful about the bigger picture” of what they’re being asked to do and whether it’s malicious or not, he added, while open-source models are “most likely to do what you say” and existing guardrails can be more easily circumvented.

The post Top AI companies have spent months working with US, UK governments on model safety appeared first on CyberScoop.

Reading this, I kept hoping that we’d find out it’s just a hoax. But alas… Kevin Collier reports: A hacker has exploited a leading artificial intelligence chatbot to conduct the most comprehensive and lucrative AI cybercriminal operation known to date, using it to do everything from find targets to write ransom notes. In a report...

Source

The Pentagon’s two-year public competition to spur the development of cyber-reasoning systems that use large language models to autonomously find and patch vulnerabilities in open-source software concluded Friday with $8.5 million awarded to three teams of security specialists at DEF CON. 

The Defense Advanced Research Project Agency’s AI Cyber Challenge seeks to address a persistent bottleneck in cybersecurity — patching vulnerabilities before they are discovered or exploited by would-be attackers. 

“We’re living in a world right now that has ancient digital scaffolding that’s holding everything up,” DARPA Director Stephen Winchell said. “A lot of the code bases, a lot of the languages, a lot of the ways we do business, and everything we’ve built on top of it has all incurred huge technical debt… It is a problem that is beyond human scale.” 

The seven semifinalists that earned their spot out of 90 teams convened at last year’s DEF CON were scored against their models’ ability to quickly, accurately and successfully identify and generate patches for synthetic vulnerabilities across 54 million lines of code. The models discovered 77% of the vulnerabilities presented in the final scoring round and patched 61% of those synthetic defects at an average speed of 45 minutes, the competition organizers said.

The models also discovered 18 real zero-day vulnerabilities, including six in the C programming language and 12 in Java codebases. The teams’ models patched none of the C codebase zero-days, but automatically patched 11 of the Java zero-days, according to the final results shared Friday.

Team Atlanta took the first-place prize of $4 million, Trail of Bits won second place and $3 million in prize money, and Theori ranked third, taking home $1.5 million. The competition’s organizers allocated an additional $1.4 million in prize money for participants who can demonstrate when their technology is deployed into critical infrastructure. 

Representatives from the three winning teams said they plan to reinvest the majority of the prize money back into research and further development of their cyber-reasoning systems or explore ways to commercialize the technology.

Four of the models developed under the competition were made available as open source Friday, and the three remaining models will be released in the coming weeks, officials said.

“Our hope is this technology will harden source code by being integrated during the development stage, the most critical point in the software lifecycle,” Andrew Carney, program manager of the competition, said during a media briefing about the challenge last week. 

Open sourcing the cyber-reasoning systems and the AI Cyber Challenge’s infrastructure should also allow others to experiment and improve upon what the competition helped foster, he said. DARPA and partners across government and the private sector involved in the program are pursuing paths to push the technology developed during the competition into open-source software communities and commercial vendors for broader adoption.

DARPA’s AI Cyber Challenge is a public-private endeavor, with Google, Microsoft, Anthropic and OpenAI each donating $350,000 in LLM credits and additional support. The initiative seeks to test AI’s ability to identify and patch vulnerabilities in open-source code of vital importance throughout critical infrastructure, including health care. 

Jim O’Neill, deputy secretary of the Department of Health and Human Services, spoke to the importance of this work during the AI Cyber Challenge presentation at DEF CON. “Health systems are among the hardest networks to secure. Unlike other industries, hospitals must maintain 24/7 uptime, and they don’t get to reboot. They rely on highly specialized, legacy devices and complex IT ecosystems,” he said. 

“As a result, patching a vulnerability in health care can take an average of 491 days, compared to 60 to 90 days in most other industries,” O’Neill added. “Many cybersecurity products, unfortunately, are security theater. We need assertive proof-of-work approaches to keep networks, hospitals and patients safer.”

Health officials and others directly involved in the AI Cyber Challenge acknowledged the problems posed by insecure software are vast, but said the results showcased from this effort provide a glimmer of hope. 

“The magnitude of the problem is so incredibly overwhelming and unreasonable that this is starting to make it so that maybe we can actually secure networks — maybe,” Jennifer Roberts, director of resilient systems at HHS’s Advanced Research Projects Agency for Health, said during a media briefing at DEF CON after the winners were announced. 

Kathleen Fisher, director of DARPA’s Information Innovation Office, shared a similar cautiously optimistic outlook. “Software runs the world, and the software that is running the world is riddled with vulnerabilities,” she said.

“We have this sense of learned helplessness, that there’s just nothing we can do about it. That’s the way software is,” she continued. The AI Cyber Challenge “points to a brighter future where software does what it’s supposed to do and nothing else.”

The post DARPA’s AI Cyber Challenge reveals winning models for automated vulnerability discovery and patching appeared first on CyberScoop.