Prompting for Patches That Fix Vulns Without Adding New Ones - Keith Hoodlet - ASW #394
Application Security Weekly (Audio)Show Notes
There's already an increase in volume of security flaws found by LLMs. And orgs are already turning to LLMs to write code. So, what happens when orgs lean on LLMs to create patches for those security flaws? Keith Hoodlet gives an exclusive early look at his team's recent research into the success, quality, and failures of LLM-generated security patches. Notably, they saw scenarios across a spectrum from robust, effective patches to patches that changed the software's behavior to patches that introduced new vulns to patches that didn't even fix the original vuln while also introducing a new vuln. The research considers factors like quality and correctness of prompts, complexity of the target software, programming language, and expertise required to understand what a robust patch should look like. If you're going to spend tokens on fixing security flaws, you want a feedback loop that fixes them correctly -- not an infinite loop of new flaws creeping in with every LLM iteration. Watch for this research, its toolset, and data to be released on Thursday August 6th during Black Hat. Visit https://www.securityweekly.com/asw for all the latest episodes! Show Notes: https://securityweekly.com/asw-394
More from Application Security Weekly (Audio)
The OWAS Mobile Application Security (MAS) project recently released an updated version of its guidance, with excellent advice organized into common weaknesses, a verification standard for devs (MASVS), and a testing guide (MASTG). Ryan Lloyd and Jan Seredynski from Guardsquare dive into the project to talk about how it's been updated, how orgs can use it, and the challenges orgs face in managing their mobile frameworks and environments. We also cover how agents and LLMs are impacting mobile security and what future iterations of the MAS project might include for securing mobile apps interacting with LLMs or being written by LLMs. Finally, we share practical ways for developers to get involved as OWASP contributors and advocates. Segment Resources: https://mas.owasp.org https://mas.owasp.org/MASWE/ https://hubs.ly/Q04ykvgm0 This segment is sponsored by Guardsquare. Visit https://securityweekly.com/guardsquare to learn more! Visit https://www.securityweekly.com/asw for all the latest episodes! Show Notes: https://securityweekly.com/asw-405
To celebrate our 404th episode we discuss what's not found in modern appsec and what still exists that should become not found. The conversation starts with a classic problem of appsec -- looking for a reliable vuln prioritization methodology. But we expand from there to challenge whether a focus on vulns and patches is even the right choice for an appsec program. Then it's on to incentives, the role of agents in writing and securing software, and what we should be looking for to bring new appsec practitioners into the industry. Visit https://www.securityweekly.com/asw for all the latest episodes! Show Notes: https://securityweekly.com/asw-404
The principle of least privilege access is ubiquitous in recommendations, but perhaps too uncommon in practice. Alex Olivier explains how the AuthZen Working Group has been approaching new standards, guidelines, and practical deployment patterns to make it easier for orgs to establish more refined and correct access policies. It's a problem that predates LLMs, but also one that agents and MCPs made all the more acute. We talk about what intent means when trying to constrain agents' activities, what role LLMs have in policy decisions, and how LLMs can help create a much-needed inventory of access controls. Segment Resources: https://www.cerbos.dev/blog/dimmer-switch-not-a-kill-switch-rethinking-ai-agent-governance https://www.cerbos.dev/blog/multi-hop-delegation-ai-agents https://openid.net/wg/authzen/ Visit https://www.securityweekly.com/asw for all the latest episodes! Show Notes: https://securityweekly.com/asw-403
Finding flaws with LLMs and agents is changing bug bounty programs. But it's not necessarily changing how orgs fix those flaws. Shlomie Liberow shares his experience across a decade of bounty programs and how they have changed for researchers and orgs. He explains why fixing the bug reported through a bug bounty is more about understanding interconnected systems than fixing a single piece of software. We discuss how orgs can be more effective at securing their environment and what role LLMs might have in evaluating controls. Plus, we look to the future of bug bounty programs and how researchers can still excel through curiosity and expertise on a topic rather than relying on scanners, prompts, and luck. Visit https://www.securityweekly.com/asw for all the latest episodes! Show Notes: https://securityweekly.com/asw-402
Prompt injection demonstrates one of the major challenges in securing LLMs and agents -- how do you ensure an agent ignores attackers and only does what you instructed it to do. The flaw highlights how LLMs mix inputs, context, and outputs without any strict boundaries between them. Julie Brunias joins us to talk through examples of injections, why their consequences can go beyond information leaks, and why trying to mitigate them with other LLMs is insufficient. Segment resources: https://llmgateway.io/open-source https://www.wiz.io/blog/off-guard-breaking-litellm-from-authentication-bypass-to-cloud-compromise Visit https://www.securityweekly.com/asw for all the latest episodes! Show Notes: https://securityweekly.com/asw-401