Ctrl+Z

Reversibility in the agentic era

1. AI's new foot soldiers could supercharge productivity—or sabotage it

While companies are spending heavily on generative AI (GenAI), the productivity impact remains muted. One recent survey found that teams primarily using GenAI were no more likely to report high productivity gains than those that did not.1 Studies from MIT2 to McKinsey3 show many are struggling to unlock hoped-for benefits. Could agents deliver? Unlike chatbots, which respond to specific prompts, agents act on a user's behalf and can even collaborate with each other. Deloitte expects a quarter of all companies using GenAI to have implemented agentic pilots by the end of 2025 on everything from updating prices to resolving IT helpdesk requests.4 The IT department is by far the biggest agentic adopter.5

Figure 1: Bot embrace

Agent adoption by function

Source: Information Services Group (2025)6

“Agentic AI is a revolution,” says Tejas Gajjar, technology manager of engineering at US retailer Macy’s. “It means AI is not really a technology anymore, but an integral part of the business, taking actions alongside human employees.”

But to operate autonomously, companies must grant agents unprecedented access to sensitive information like payment processes, databases and passwords. “The business model of AI designers is to reduce human control in order to have fluid, automated workflows,” argues Vivien Mura, chief technology officer at Orange Cyberdefense, a cybersecurity services subsidiary of Orange Group. This handover creates novel cyber risks.

“Last year companies were in a race to adopt AI, so security was not the priority. Now, a lot of customers are beginning to ask us for the best manner to secure their AI.”

Vivien Mura, chief technology officer at Orange Cyberdefense

2. Fueling the cyber wildfire

The cybersecurity environment has deteriorated significantly. Confirmed breaches have quadrupled in the last six years, according to data from Verizon.7

“Agents can transfer ideas from one domain to another very efficiently. Imagine having agents that can manipulate models and transport ideas. In principle, they will find new solutions [and] new forms of attack that we don’t even know of yet.”

Dr Boris Bolliet, agentic AI research lead at the Infosys-Cambridge AI Centre

Could the rush to deploy agents make things worse? Malicious actors can extract sensitive information directly from an agent or feed it data to influence decisions in nefarious ways. "AI technologies are insecure by default," says Mr Mura at Orange Cyberdefense. "There are a lot of vulnerabilities that will be inserted, both at the software level and in AI models." On the dark web, he says, hackers are sharing tactics for harnessing agents. Agents also improve cybercriminals’ productivity, helping them orchestrate more attacks.

Dr Boris Bolliet, agentic AI research lead at the Infosys-Cambridge AI Centre, worries that the same ingenuity shown by agents in areas like scientific discovery, his field of study, is coming to the cybersphere.

Agents can wreak havoc through errors too. Eight in ten companies in one study said their AI agents have taken unintended actions; 39% reported that agents had accessed unauthorized systems or resources; and 33% found that agents had inadvertently shared sensitive data.8 In one infamous case, an agent deleted a company's entire codebase.9 Elsewhere, a customer support agent for a large tech company leaked a customer relationship manager (CRM) database.10 Security researchers have also found voice assistants feeding sensitive data back to tech companies.11 The outage of CrowdStrike’s Falcon platform, which automatically monitors computer systems for cyber threats, caused worldwide shutdowns from hospitals to airports, showing how losses from automation slipups can be as or more severe than cybercrime attacks.12

The impacts go far beyond the IT department. The global average cost of a breach is US$4.44m, according to IBM.13 According to a 2024 study, the vast majority (90%) of large and mid-size enterprises reported that a single hour of downtime costs upwards of $300,000 on average.14 Publicly traded companies suffered a 7.5% average decline in their stock values after a data breach and a mean market cap loss of US$5.4 billion.15 Companies with weaker cybersecurity practices can face higher borrowing costs and increased financial risk. Moody’s reduced Equifax’s credit rating in 2019 following a data breach that the company experienced in 2017.16 Regulatory punishments are growing too. The EU’s Digital Operational Resilience Act (DORA) requires financial services industries to be able to respond and recover effectively to all digital disruptions.17

This convergence of intensifying cyber threats and AI’s intrinsic vulnerabilities will force companies to augment existing preventive strategies with a focus on rapid recovery and ‘reversibility’.

3. Hit rewind

The acceleration of cybersecurity risk in the AI era is changing organizational security strategy, with the ability to respond and react now as critical as prevention. A study of 3.6 billion observed cyber vulnerabilities found that over half were still present a month after discovery and nearly a fifth remained after a year (see Figure 2). It took more than a month for Marks & Spencer, a retailer, to recover from its recent hacks.18 Unresolved vulnerabilities must be more speedily resolved in the agentic era.

Figure 2: Don’t let it linger

Speed of response to vulnerabilities

Source: Cyentia and Cisco (2025)19

As agents are embedded into enterprise technology, organizations will need full transparency of their activities via an inventory of agents, an indication of decisions they are empowered to take and the sensitive information they access. This allows vulnerabilities to be monitored and powers withdrawn when an agent no longer needs them. Thankfully, this observability might be easier with agents than humans, who remain the leading source of data breaches (see Figure 3).

Figure 3: The human factor

Key enumerations in breaches

Source: Verizon (2025)20

Observability is only the first step. Enterprises need tools to allow IT teams to rewind to any point in a process and guarantee other systems will reflect this reset. While companies develop such tools, AI agents should be limited to very specific functions in controlled pilots. If companies invest in tools and systems for undoing actions, the programmatic nature of agents lends itself to reversibility. AI leaders are also investing engineering resources to improve agentic safety. Anthropic has developed tools that can govern the interaction of agents and provide safeguards to their actions.21 Google DeepMind tests conversational agents to expose and weed out rule breaches.22

Lastly, companies should look to agents as tools of defense. The retail industry, for example, struggles to maintain operational IT stability during peak periods like the lead-up to holidays such as Thanksgiving. “Agents can identify risks based on predictive analytics and make sure patches [vulnerability fixes] are rolled out,” notes Mr Gajjar at Macy’s. According to ISG, nearly 10% of the IT use cases were cyber-related.23

Companies will not fund AI pilots indefinitely. Soon, they will need to deliver more, and productivity will be a leading success indicator. Agents could be the difference between AI as a helpful query assistant to a self-directed engine of activity. That will add security risks as powers and access are handed over. On top of existing cybersecurity threats, organizations will need to invest as much in reversibility of the inevitable breaches as they have in prevention.

"Everything AI agents do is being recorded. I can look back and understand what has been done."

Dr Boris Bolliet, agentic AI research lead at the Infosys-Cambridge AI Centre

References