From advanced cyber‑attacks to the rise of deepfake porn, political misinformation, and chatbot‑induced psychosis, the darker edge of AI seems to be evolving quicker than humans can build the safeguards to contain it. 

An extensive new AI safety study, penned by over 100 AI experts including Turing Award-winner Yoshua Bengio – one of AI’s many “godfathers” – documents a striking, near-terrifying spike in real-world AI malfunctions, misuse, and systemic risks to humanity.

Designed to help global policymakers navigate an ‘evidence dilemma’ in which these risks only gradually surface relative to AI’s capabilities, the International AI Safety Report 2026 sets out exactly where the biggest dangers lie – and what, if anything, we can do about them.

Among the most prominent risks highlighted in the report are those already playing out in the real world, especially when it comes to the use of general-purpose AI to create harmful content, but current policies are struggling to keep pace.

For example, while the controversy around Grok’s sexualised deepfakes highlights the personal harm AI misuse can inflict, the report warns that evidence of the wider impact remains thin, as underreporting and poor detection tools make it hard to track these harms.

There is also the growing misuse of AI by hackers, with security researchers routinely observing malicious actors using AI, with the report noting that AI systems are now particularly good at rooting out software vulnerabilities and writing malicious code.

Last year, these capabilities saw an AI agent in a cyber competition uncover 77% of vulnerabilities in real software, outperforming all but the top 5% of more than 400 (mostly human) teams.

Again, though, tracking the evolution and proliferation is extremely difficult for governments and law enforcement, especially with underground marketplaces now selling pre-packed AI tools that have lowered the bar for entry.

There is also the issue of AI agents growing impact on the physical world, with the study warning that their ability to take independent, unobserved actions introduces risks we can’t yet conceive. 

That is troubling, given that critical sectors like financial services and energy management are pulling further ahead in autonomous AI development. 

Recommended reading

• 69% of C-Suite Leaders Choose Fast AI Over Secure AI
• Can AI De‑Risk Breakneck Data Centre Builds?
• What Will AI Sovereignty Look Like?

While those issues are well known, if tricky to solve, the report goes further in its doom-mongering, warning that several AI firms tightened safeguards last year after tests revealed their systems could, in theory, help develop biological and chemical weapons. 

Although the report stresses that AI is unlikely to help wannabe terrorists develop such weapons in the near term, the fact that it lowers expertise barriers may one day allow such malicious actors to cause immeasurable harm.   

Despite all that nightmare fuel, however, the study stresses there is still time to get ahead of these problems, starting with developers, governments, and institutions sharing evidence and collaborating on a common foundation of AI safety.

At the moment, that is unimaginable, given the intense international competition around the technology.

“Since the release of the inaugural International AI Safety Report a year ago, we have seen significant leaps in model capabilities, but also in their potential risks, and the gap between the pace of technological advancement and our ability to implement effective safeguards remains a critical challenge,” said Yoshua Bengio, scientific advisor of the Mila – Quebec AI Institute.

“The Report is intended to provide decision-makers with the rigorous evidence needed to steer AI toward a future that is safe, secure, and beneficial for all.”