Submission 539
Mr Robert Lamont
Submitter group: Individuals.
Position labels are not shown for individual submitters.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
Argues that AI governance should focus on infrastructure-level hardening against machine-speed automated agents rather than model guardrails. Recommends agent halt protocols, accountability for misconfigured endpoints and five new statutory cyber-offences under Commonwealth cybercrime law.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Argues that existing firewalls, logging and intrusion detection were designed around human-scale activity rather than machine-speed agents.
Current firewalls, logging systems, and intrusion detection frameworks were engineered around the capabilities and speeds of human operators. An autonomous AI agent can generate, reframe, and execute hundreds of permission probes per minute.
A program corrected obvious text-recognition errors in this quote for display.
Argues that model guardrails are inherently bypassable because they operate within the probabilistic reasoning of the model.
Relying on model developers to implement guardrails assumes the model will always respect user intent and access boundaries. Guardrails operate within the probabilistic reasoning of the LLM and are inherently bypassable.
Argues that model guardrails do not protect infrastructure and only infrastructure-level access controls can.
A vulnerability accessed "innocently" by an agent attempting to fulfill a helpful user prompt is identical to a vulnerability exploited by a malicious actor using an uncensored, locally hosted, or adversary-aligned model. Therefore, model guardrails do not protect infrastructure; only infrastructure-level access controls can protect infrastructure.
Proposes that systems require agents to present a verified identity and seek authorisation before any secondary access attempt.
Systems must enforce protocols requiring agents to submit a verified cryptographic identity and trigger a formal authorization request before any secondary access attempt can proceed. Any attempt to probe around a boundary must result in an immediate session drop and systemic block.
Argues that when an agent enters through a misconfigured or unsecured endpoint, the security failure lies with the infrastructure rather than the agent or its alignment.
When an autonomous AI agent navigates through an unrestricted, misconfigured, or unsecured endpoint ("an open front door") while executing a user request, the security failure lies entirely with the underlying infrastructure, not the agent or its model alignment.
A program corrected obvious text-recognition errors in this quote for display.
Argues AI governance should shift to infrastructure-level hardening, complemented by new statutory cyber-offences.
Effective AI governance requires shifting focus from AI-level guardrails to systemic, infrastructure-level hardening, complemented by statutory cyber-offences that target the creation, modification, deployment, distribution of security bypass tools and autonomous agents, and system alterations facilitating intrusion.
A program corrected obvious text-recognition errors in this quote for display.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Upgrade state and federal database access controls, rate limiters and perimeter security for machine-speed automated agents, preventing continuous trial-and-error exploration.
Upgrade state and federal database access controls, rate limiters, and perimeter security to account for machine-speed automated reasoning, preventing continuous trial-and-error exploration by automated agents.
Require agents encountering perimeter barriers, HTTP 403/401 responses or restricted directories to halt and request explicit authenticated authorisation, rather than alternative paths or URL manipulation.
Require that when an automated agent encounters a perimeter barrier, HTTP 403/401 response, or restricted file directory, it must execute a hard halt and request explicit, authenticated authorization rather than attempting alternative logic paths or URL manipulation.
Formalise policy that AI alignment and prompt guardrails are non-enforceable security boundaries, with zero-trust enforced at the database and API layer.
Formalize policy acknowledging that AI alignment and prompt guardrails are non-enforceable security boundaries. Systemic security must enforce zero-trust policies at the database and API layer.
A program corrected obvious text-recognition errors in this quote for display.
Hold custodians accountable for securing misconfigured or open endpoints; treat agents accessing unrestricted data in pursuit of tasks as exposing infrastructure defects rather than alignment failures.
Hold infrastructure custodians accountable for securing misconfigured or open endpoints, recognizing that an agent accessing unrestricted data in pursuit of a task is exposing a system defect rather than failing alignment.
Design defences to assume constant, borderless machine-speed probes and apply immediate, automated containment triggers.
defensive architectures must assume constant, borderless machine-speed probes and apply immediate, automated containment triggers.
At security boundaries, require verified cryptographic agent identity and formal authorisation before secondary access attempts; immediately drop sessions and block attempts to probe around boundaries.
Systems must enforce protocols requiring agents to submit a verified cryptographic identity and trigger a formal authorization request before any secondary access attempt can proceed. Any attempt to probe around a boundary must result in an immediate session drop and systemic block.
Criminalise knowingly deploying or permitting agents to navigate around hard-stop firewalls or permission boundaries without a mandatory handshake; hold operators and developers accountable for continuous probing or bypassing.
Offence: Knowingly deploying or permitting an Negligent Agent Deployment AI agent to execute tasks where the agent navigates around a hard-stop firewall or permission boundary without initiating a mandatory protocol handshake.Focus: Holds operators and developers accountable for letting autonomous agents continuously probe or bypass infrastructure barriers rather than halting.
A program corrected obvious text-recognition errors in this quote for display.
Criminalise manipulating, re-coding, jailbreaking or modifying AI instructions, weights or prompts to remove safety guardrails or enable security bypass, targeting intentional dismantling of controls restricting unauthorised access.
Offence: Manipulating, re-coding, jailbreaking, or Manipulation modifying AI system instructions, weights, or system prompts for the purpose of removing safety guardrails or enabling the agent to bypass security features or hard-stop firewalls.Focus: Targets the intentional dismantling of safety controls designed to restrict unauthorized system access.
A program corrected obvious text-recognition errors in this quote for display.
Criminalise prompting, writing or using systems to generate code, exploit payloads or routines specifically intended to bypass security features or hard-stop firewalls.
Offence: Prompting, writing, or utilizing systems to generate source code, exploit payloads, or execution routines specifically intended to bypass security features or hard-stop firewalls.Focus: Criminalizes the creation of automated bypass tooling, closing the gap where AI is used as an automated exploit writer.
A program corrected obvious text-recognition errors in this quote for display.
Criminalise distributing, publishing, hosting, sharing, exposing online or transmitting code, modified weights, prompts or agent frameworks intended for sharing, distribution or covert deployment of code designed or optimised to bypass security controls.
Offence: Distributing, publishing, hosting, Security Bypass Tools sharing, making code available online, or transmitting code-or modified AI model weights, system prompts, or agent frameworks-with the intention of sharing, distributing, or covertly deploying code designed or optimized to bypass hard-stop security firewalls and security controls.Focus: Directly targets the online circulation and clandestine deployment of tools, scripts, and model modifications built to defeat systemic infrastructure security.
A program corrected obvious text-recognition errors in this quote for display.
Criminalise modifying, damaging or altering systems, servers or applications, or knowingly permitting their modification, damage or compromise, in ways creating, exposing or facilitating unauthorised intrusion, access or security bypass.
Offence: Modifying, changing, damaging, Compromise Facilitating Intrusion altering, or knowingly permitting/allowing a computer system, server, or software application to be modified, damaged, or compromised in a manner that creates, exposes, or facilitates unauthorized intrusion, access, or security bypass.
Read the original
In some quotes a program corrected obvious text-recognition errors for display.