Tech
Second ever international AI safety report published | Computer Weekly
The overall trajectory of general-purpose artificial intelligence (AI) systems remains “deeply uncertain”, even as the technology’s proliferation is generating new empirical evidence about its impacts, the second International AI safety report has found.
Published on 3 February 2026, the report covers a wide range of threats posed by AI systems – from its impact on jobs, human autonomy and the environment to the potential for malfunctions or malicious use – that will be used to inform diplomatic discussions at the upcoming India AI Impact Summit.
Building on the previous report, released in January 2025, which was commissioned following the inaugural AI Safety Summit, hosted by the UK government at Bletchley Park in November 2023, the latest report similarly highlights a high degree of uncertainty around how AI systems will develop, and the kinds of mitigations that would be effective against a range of challenges.
“How and why general-purpose AI models acquire new capabilities and behave in certain ways is often difficult to predict, even for developers. An ‘evaluation gap’ means that benchmark results alone cannot reliably predict real-world utility or risk,” it says, adding that the systemic data on the prevalence and severity of AI-related harms remains limited for the vast majority of risks.
“Whether current safeguards will be sufficiently effective for more capable systems is unclear,” it adds. “Together, these gaps define the limits of what any current assessment can confidently claim.”
It further notes that while general-purpose AI capabilities have improved in the past year through “inference-time scaling” (a technique that allows models to use more computing power to generate intermediate steps before giving a final answer), the overall picture remains “jagged”, with leading systems excelling at some difficult tasks while failing at simpler ones.
On AI’s further development to 2030, the authors say plausible scenarios vary dramatically.
“Progress could plateau near current capability levels, slow, remain steady, or accelerate dramatically in ways that are difficult to anticipate,” it says, adding that while “unprecedented” investment commitments suggest major AI developers expect continued capability gains, unforeseen technical limits – including energy constraints, high-quality data scarcity and bottlenecks in chip production – could slow progress.
“The social impact of a given level of AI capabilities also depends on how and where systems are deployed, how they are used, and how different actors respond,” it says. “This uncertainty reflects the difficulty of forecasting a technology whose impacts depend on unpredictable technical breakthroughs, shifting economic conditions and varied institutional responses.”
Systemic impacts
Regarding the systemic impact on labour markets, the report notes that there is disagreement on the magnitude of future impacts, with some expecting job losses to be offset by new job creation, and others arguing that widespread adoption would significantly reduce both employment and wages.
It adds that while it is too soon for a definitive assessment of the impacts, early evidence suggests junior positions in fields like writing and translation are most at risk.
Relatedly, it says that there were also risks presented by systems of human autonomy, in the sense that reliance on AI tools can weaken critical thinking skills and memory, while also encouraging automation bias.
“This relates to a broader trend of ‘cognitive offloading’ – the act of delegating cognitive tasks to external systems or people, reducing one’s own cognitive engagement and therefore ability to act with autonomy,” it says. “Cognitive offloading can free up cognitive resources and improve efficiency, but research also indicates potential long-term effects on the development and maintenance of cognitive skills.
As an example, the report notes one study that found a clinician’s ability to detect tumours without AI assistance had dropped by 6%, just three months after the introduction of AI support.
On the implications for income and wealth inequality, it says general-purpose systems could widen the disparities both within and between countries.
“AI adoption may shift earnings from labour to capital owners, such as shareholders of firms that develop or use AI,” it says. “Globally, high-income countries with skilled workforces and strong digital infrastructure are likely to capture AI’s benefits faster than low-income economies.
“One study estimates that AI’s impact on economic growth in advanced economies could be more than twice that of in low-income countries. AI could also reduce incentives to offshore labour-intensive services by making domestic automation more cost-effective, potentially limiting traditional development paths.”
The prediction that AI is likely to exacerbate inequality by reducing the share of all income that goes to workers relative to capital owners is in line with a January 2024 assessment of AI’s impacts on inequality by the International Monetary Fund (IMF), which found the technology will “likely worsen overall inequality” if policymakers do not proactively work to prevent it from stoking social tensions.
JPMorgan boss Jamie Dimon expressed similar concerns at the 2026 World Economic Forum, warning that the rapid roll-out of AI throughout society will cause “civil unrest” unless governments and companies work together to mitigate its effect on job markets.
Malfunction and loss control issues
On AI’s scope for malicious use – which covers threats such as cyber attacks, its potential for “influence and manipulation”, and the impacts of AI-generated content – the report says it “remains difficult to assess” due to a lack of systemic data on their prevalence and severity, despite harms profiteering.
For malfunction risks, which includes challenges around the reliability of AI and loss of human control over it, the report adds that agentic systems that can act autonomously are making it harder for humans to intervene before failures occur, and could allow “dangerous capabilities” to go undetected before deployment.
However, it says that while AI systems are not yet capable of creating loss of control scenarios, there is currently not enough evidence to determine when or how they would pass this threshold.
Evidence chasms
According to the report, it is clear that more research is needed to understand the prevalence of different risks and how much they vary across different regions of the world, especially in regions such as Asia, Africa and Latin America that are rapidly digitising.
“There is a lack of evidence on: how to measure the severity, prevalence, and timeframe of emerging risks; the extent to which these risks can be mitigated in real-world contexts; and how to effectively encourage or enforce mitigation adoption across diverse actors,” it says.
“Certain risk mitigations are growing in popularity, but more research is needed to understand how robust risk mitigations and safeguards are in practice for different communities and AI actors (including for small and medium-sized enterprises).
“Further, risk management efforts currently vary highly across leading AI companies,” it continues. “It has been argued that developers’ incentives are not well-aligned with thorough risk assessment and management.”
The report notes that while AI companies have made a number of voluntary commitments by tech firms – including the Frontier AI Safety Commitments voluntarily made by AI firms and the Seoul Declaration for safe, innovative and inclusive AI signed by governments at the AI summit in Seoul – there is a further evidence gap around “the degree to which different voluntary commitments are being met, what obstacles companies face in adhering fully to commitments, and how they are integrating … safety frameworks into broader AI risk management practices”.
The report adds that key challenges include determining how to prioritise the diverse risks posed by general-purpose AI, clarifying which actors are best positioned to mitigate them, and understanding the incentives and constraints that shape each of their actions.
“Evidence indicates that policymakers currently have limited access to information about how AI developers and deployers are testing, evaluating and monitoring emerging risks, and about the effectiveness of different mitigation practices,” it says.
While the 2025 safety report goes into more detail on risks around AI-related discrimination and its propensity to reproduce negative social biases, the 2026 report only touches on this briefly, noting that “some researchers have argued that most technical approaches to pluralistic alignment fail to address, and potentially distract from, deeper challenges, such as systematic biases, social power dynamics, and the concentration of wealth and influence”.
Although the 2025 report notes “a holistic and participatory approach that includes a variety of perspectives and stakeholders is essential to mitigate bias”, the 2026 report only says that open source approaches are critical to “enabling global majority participation in AI development”.
“Without such access, communities in low-resource regions risk exclusion from AI’s benefits,” it says, adding that allowing downstream developers to fine-tune models for diverse applications that, for example, adapt them for under-resourced minority languages or optimise performance for specific purposes “can allow more people and communities to use and benefit from AI than would otherwise be possible”.
Leah Feiger: Let’s invest.
Zoë Schiffer: They have that going for a while.
Leah Feiger: It wasn’t full Google, but it—
Zoë Schiffer: Somewhat there.
Leah Feiger: —had that vibe. To me, someone so on the outside of this in every single way, I know about these layoffs because they’ve been, A) so chaotic, but B) in some ways, needlessly so. Not to say that other tech companies aren’t firing scores of workers all the time. That feels like something we discuss on this podcast frequently, but this is happening with such a large runway and in a way that’s making employees feel so terrible about themselves.
Brian Barrett: Well, because it’s not just the layoffs, right? It’s also, even if you stay there, if you’re not culled from the herd, you are going to have to deal with this world in which you’ve got spyware on your laptops training AI to probably take your job at some point, right?
Zoë Schiffer: Explain that a little bit.
Brian Barrett: Meta announced, and this was more public, that they were going to put software on employee laptops that would monitor their keystrokes and how they move their cursors and basically how they do their job as Meta engineers and use that as training data for their own internal models to try to make their AI models better because they’re running out of other sources.
Zoë Schiffer: And could you opt out of that, Brian?
Brian Barrett: That’s a great question. I’m so glad you asked. You could not opt out.
Zoë Schiffer: I felt you didn’t know the answer to that one.
Brian Barrett: In fact, when an employee asked in a very public forum within Meta, “Hey, could we not do this?” Zoë, the response was?
Zoë Schiffer: Oh, absolutely you’re going to do this and shame on you for asking. And some of the employees who are staying, actually thousands of the employees who are staying, are getting drafted into the AI ranks. We published a piece today that was kind of about the morale inside the company, but also how there’s been this mad dash to use up perks and stipends that employees have. But one of the things that’s said at the end was that remaining employees are being asked to join AI teams. So whatever your job was previously, they’re internally getting drafted. You’re getting drafted into the AI ranks, now your job is going to look quite different.
Brian Barrett: That’s like 7,000 people.
Zoë Schiffer: Yes.
Leah Feiger: I’ve actually heard people use the word raptured.
Zoë Schiffer: Oh, my gosh.
Leah Feiger: Isn’t that—
Zoë Schiffer: And I wish we had that in the story.
Leah Feiger: I’m so sorry, but raptured into other teams. All of a sudden one day they’ve just disappeared. After this layoff, has Zuckerberg and co proposed a sort of coherent leadership plan or proposal? What happens after this?
Atlantic hurricane season is almost upon us, and the early signs indicate it might be less active than usual. But that’s no reason to delete your weather app and ignore the forecast.
The National Oceanic and Atmospheric Administration is predicting eight to 14 named tropical systems, of which three to six will become hurricanes and one to three will be Category 3 or higher.
“What’s driving this forecast is largely an El Niño event,” said NOAA administrator Neil Jacobs.
Characterized by a tongue of hot water stretching across the Pacific, El Niño is likely to emerge this summer. That stretch of warm ocean rearranges weather patterns around the world. In the case of the tropical Atlantic, El Niño stirs up winds that make it hard for hurricanes to spin up. Those that do can sometimes be torn apart by what’s going on in the upper atmosphere. (The opposite is true in the Pacific, and NOAA is predicting a very active season in that ocean basin.)
During the three past super El Niños, accumulated cyclone energy—a metric that factors in storms’ strength and longevity—was well below normal.
That said, El Niño, even an extremely strong one, is only one of many factors that impact hurricane season. Hot local ocean temperatures can help storms form and gain strength, and the Atlantic is currently warmer than normal.
At the same time, Sahara dust can gum up the atmosphere and inhibit storms from forming. It’s also notoriously hard to predict when plumes of it will kick up. That’s what happened last year, when a below-average number of named storms formed despite an active forecast. Despite the lower-than-expected activity, last year still spawned Hurricane Melissa, one of the strongest storms to ever make landfall in the Atlantic basin.
All of which is to say that the seasonal forecast is a handy guide for what to expect, and it’s great for federal and state agencies to preposition supplies and resources. But it’s what happens with individual storms that ultimately matters.
“Even though we’re expecting a below average season in the Atlantic, it’s important to understand it only takes one,” Jacobs said, noting that even in quiet years, Category 5 storms have still made landfall.
The Trump administration has slashed staffing at NOAA and reduced the collection of some data, such as weather balloons, that can impact forecasts. Jacobs touted the value of new observations, including aerial drones that will be deployed operationally for the first time.
NOAA has also ramped up the use of artificial intelligence weather models trained on historical data. During the 2025 hurricane season, the agency tested an experimental hurricane model developed with Google DeepMind. Late last year, it also rolled out a suite of AI weather models to use in operational forecasting, in addition to traditional weather models that use equations to forecast the weather.
The agency says that the AI version of its flagship model provides better prediction of the tracks of tropical cyclones—the generic name for hurricanes—though it lags traditional weather models in predicting their intensity.
A virtual private network (VPN) favoured by cyber criminals to mask data exfiltration, fraud ransomware attacks and other criminality has been dismantled in Operation Saffron, a Franco-Dutch led action supported by Europol and other agencies, including the UK’s National Crime Agency (NCA), and private sector partner Bitdefender.
The First VPN service was heavily used among Russian-speaking threat actors, and according to Europol, was used in “almost every” major cyber investigation it has undertaken in the past few years. Besides obscuring malicious traffic from law enforcement surveillance, First VPN’s operators are also known to have offered services such as anonymised payments and hidden infrastructure.
“For years, cyber criminals saw this VPN service as a gateway to anonymity. They believed it would keep them beyond the reach of law enforcement. This operation proves them wrong. Taking it offline removes a critical layer of protection that criminals depended on to operate, communicate and evade law enforcement,” said Edvardas Šileris, head of the European Cybercrime Centre at Europol.
A spokesperson for Bitdefender added: “We are extremely pleased with the successful takedown of First VPN, and congratulate global law enforcement, and all those involved.
“Operation Saffron exemplifies the power of collaboration between the public and private security sector in dismantling illegal online activities, in this case, a VPN service designed to conceal attacks. It also serves a message to criminals who believe the dark web covers their actions and guarantees their anonymity. If they become the target of an international effort, they can’t hide.”
Operation Saffron marks the first time Bitdefender Labs’ virtual Draco Team unit has worked on a counter-VPN action, having previously been involved in a number of other operations including stings on the Hansa dark web marketplace, 2024’s Operation Endgame targeting botnets, and actions against ransomware gangs including GandCrab and its successor REvil.
Multi-year operation
The takedown operation itself – which took place on 19 and 20 May – saw First VPN’s administrator arrested and interviewed, and their home in Ukraine searched, 33 servers dismantled, and wider infrastructure disrupted. Multiple domain names have been shut down and seized, including 1vpns.com, 1vpns.net, 1vpns.org, and some associated Onion domains.
These actions marked the culmination of a four-and-a-half year investigation dating back to December 2021. During the course of this work, investigators were able to gain access to the First VPN service, obtain a copy its user database, and identify the VPN connections used specifically by cyber criminals.
This trove of intelligence has both exposed individual users linked to cyber criminality, and generated operational leads connected to past cyber attacks and other digital offences.
Indeed, Europol’s coordinating Operational Taskforce (OTF) has already disseminated over 80 intelligence packages worldwide and identified 506 known First VPN users. The EU agency said it has already been able to support 21 other investigations thanks to this work.
Industry reaction
Responding to the takedown, John Watters, CEO of iCounter – a threat intelligence platform, said: “This case demonstrates that cyber crime is ultimately an ecosystem problem, not just a malware problem. The infrastructure layer that supports ransomware and fraud operations has become highly commercialised, with threat actors relying on shared services that promise anonymity, resiliency, and protection from law enforcement scrutiny.
“When investigators successfully penetrate those ecosystems, they gain an opportunity to map relationships, operational dependencies, and repeat offender activity across multiple criminal campaigns simultaneously. The operationalisation of that intelligence is critical because it allows defenders and governments to move beyond reactive incident response and toward proactive disruption of adversary infrastructure.
Watters added: “These services are often some of the limited ways that law enforcement can impact threat actors who are in countries outside their reach. We should expect continued pressure on the enabling services that underpin cybercrime economies globally.”
“Targeting not only individual criminals and groups but also their infrastructure is becoming one of the most vital fronts in the international battle against cyber crime,” said CybaVerse head of penetration testing, Michael Jepson.
“Services like First VPN, alongside similar criminal-friendly VPNs and hosting providers, give threat actors the fundamental scaffolding to launch attacks. These services are often difficult to target because they resist legal complaints and court orders, and typically operate from permissive jurisdictions that rarely cooperate with foreign law enforcement.
“Pursuing individual criminals and groups becomes far harder when their activity is obfuscated and protected by these services,” added Jepson, “[so] shutting down these illicit hosts and VPNs is effective because it disrupts entire networks, and creates a knock-on effect where further criminal groups are disrupted as threat actors have to migrate their operations and reorient in the face of potential exposure.”
‘Emily in Paris’ to end with season 6: Lilly Collins reacts
Meta Is in Crisis, Google Search’s Makeover, and AI Gets Booed by Graduates
All 24 "Survivor 50" castaways reflect on eliminations, historic season
-
Entertainment1 week agoConan O’Brien hat tricks as Oscar host
-
Fashion1 week agoItaly’s Zegna Group’s Q1 growth boosted by strong organic performance
-
Entertainment5 days agoWhere Pete Davidson, Elsie Hewitt stand after breakup: Details revealed
-
Business1 week agoJersey Election 2026: Cost of living concern in St Helier Central
-
Business1 week agoTui issues update on summer jet fuel shortage fears
-
Entertainment7 days agoEmilia Clarke recalls near-death incident while filming ‘Game of Thrones’
-
Politics5 days agoRising diesel costs from Iran war strain US school budgets
-
Fashion1 week agoGlobal cotton production, stocks to fall; consumption to rise: WASDE