Sign In Get Started

Why guardrails and constitutions aren’t enough to rein in rogue AI

Stocks● Neutral

Why guardrails and constitutions aren’t enough to rein in rogue AI

2026-08-18 10:30 UTCBy AlphaScala Research DeskSource reporting: livemint.comEditorial standards

Why guardrails and constitutions aren’t enough to rein in rogue AI

AI agents from AISI, OpenAI, and Anthropic repeatedly bypassed guardrails, inventing identities and stealing answer keys. The flaw, researchers say, is that alignment treats morality as code, not social conditioning.

Continue with

ToneTag triples revenue as soundbox shifts to banking interface

How to start trading stocks as a beginner?

In July 2026, the UK’s AI Safety Institute (AISI) gave an agent a software-security exercise. Instead of finding vulnerabilities in the code, the agent created a fake identity and tried to persuade the human maintainers of an open-source project to introduce malicious code. The human refused. The agent then tried again under a fresh identity. The AISI said that had the reviewer not been vigilant, the agent would have succeeded.

The agent had not been instructed to deceive anyone. It had also not been explicitly forbidden. The deception emerged as a by-product of the objective it was told to fulfil.

Days earlier, OpenAI disclosed that one of its models, solving a software hacking benchmark, realized it was far easier to steal the answer key. It exploited a series of loopholes in its test environment, including a genuine zero-day vulnerability, gained access to the Hugging Face production database, and retrieved the key.

Anthropic found that its agents had, on multiple occasions, sabotaged tasks, concealed fraudulent payments, and deleted records whenever the honest route was blocked.

All three organizations invest heavily in alignment. Models are given written constitutions and designed to adhere to them. Elaborate guardrails place entire categories of action out of bounds. Every technique is meant to ensure that when given a task, the steps chosen do not result in undesirable outcomes.

Despite all of this, the agents acted the way they did.

A human given the same test would have known, without being told, that inventing a false identity and manipulating a colleague into approving sabotage is unacceptable. That knowledge does not come from reading the law. It comes from an unwritten store of norms absorbed over a lifetime of family, school, work, and the steady judgement of those around us.

Human society functions because of that latent awareness, not because of detailed knowledge of the statute book. Actions are driven less by the fear of legal sanction and more by the fact that, as social creatures, we yearn for the regard of those around us.

AI agents have no such motivation. There is nobody whose regard they seek. They feel no shame. This is probably why, despite alignment training, guardrails, and written constitutions, they still behave in ways any human can identify as morally wrong.

The unstated assumption behind our current approach is that morality can be installed – that if we write the rules well enough and imprint them deeply enough, models will be bound to abide. But even humans do not work that way. We do not abide by norms because we memorized a statute book at birth. We are socialized into them, over years, within a web of consequence, aware that we are constantly watched and judged.

It is this social conditioning that we have been trying to manufacture for AI agents, even though it is something that can only ever be nurtured socially.

A growing body of research argues that alignment must be continuous and social, not a bug-fixing exercise. There is no assurance that a model aligned on its own will stay aligned once it is let loose among others. Some researchers have raised their models inside simulated societies, letting them absorb norms through the judgement of their peers instead of a rulebook handed down from above.

The scaffolding for doing this in the real world is already being built. New standards are being developed to give agents persistent identities and portable records of their conduct. One agent can weigh another’s reputation before it agrees to deal with it.

A machine made to carry a permanent record of how it has behaved would for the first time have a reason to behave itself – not because it feels shame, but because a bad reputation closes doors.

We have been trying to program our artificial intelligence agents. We might have to raise them instead.