Coordinating Attack Response at Internet Scale 2 (CARIS2) Workshop Report

Workshop Collaboration Both CARIS workshops brought together a set of individuals who had not previously collaborated toward the goals of scaling attack response. This is important as the participants span various areas of Internet technology work, conduct research, provide a global perspective, have access to varying data sets and infrastructure, and are influential in their area of expertise. The specific goals, contributions, and participants of the CARIS2 workshop were all considered in the design of the breakout sessions to both identify and advance research through collaboration. The breakout sessions varied in format to keep attendees engaged and collaborating; some involved the full set of attendees while others utilized groups. The workshop focused on identifying potential areas for collaboration and advancing research.

Standardization and Adoption: identify widely adopted and pervasive standard protocols and data formats as well as those that failed.
Preventative Protocols and Scaling Defense: identify protocols to address automation at scale.
Incident Response Coordination: brainstorm what potential areas of research or future workshops could be held to improve on the scalability of incident response.
Monitoring and Measurement: brainstorm methods to perform monitoring and measurement with the heightened need and requirement to address privacy.
Taxonomy and Gaps: brainstorm a way forward for the proposed SMART Research Group.

Breakout 1 Results: Standardization and Adoption This breakout session considered points raised in the preceding talks on hurdles for automating security controls, detection, and response; the teams presenting noted several challenges they still face today. The breakout session worked toward identifying standard protocols and data formats that succeeded in achieving adoption as well as several that failed or only achieved limited adoption. The results from the evaluation were interesting and could aid in achieving greater adoption when new work areas are developed. The following subsections detail the results.

Wide Adoption The Transport Layer Security (TLS) protocol has replaced the Secure Sockets Layer (SSL) protocol. Observations: There was a clear need for session encryption at the transport layer to protect application data. E-commerce was a driving force at the time with a downside to those who did not adopt. Other positive attributes that aided adoption were modular design, clean interfaces, and being first to market. The Simple Network Management Protocol (SNMP) enables configuration management of devices with extension points for private configuration and management settings. SNMP is widely adopted and is only now, after decades, being replaced by a newer alternative, YANG (a data modeling language) that facilitates configuration management via the Network Configuration Protocol (NETCONF) or RESTCONF. SNMP facilitated an answer to a needed problem set: configuration, telemetry, and network management. Its development considered the connection between the user, vendor, and developers. Challenges did surface for adoption from SNMPv1.1 to 1.2, as there was no compelling reason for adoption. SNMPv3 gained adoption due to its resilience to attacks by providing protection through improved authentication and encryption. IP Flow Information Export (IPFIX) was identified as achieving wide adoption for several reasons. The low cost of entry, wide vendor support, diverse user base, and wide set of use cases spanning multiple technology areas were some of the key drivers cited. X.509 was explored for its success in gaining adoption. The solution being abstract from crypto, open, customizable, and extensible were some of the reasons cited for its successful adoption. The team deemed it a good solution to a good problem and observed that government adoption aided its success.

Limited Adoption Next, each team evaluated solutions that have not enjoyed wide adoption. Although Structured Threat Information eXpression (STIX) and the Incident Object Description Exchange Format (IODEF) are somewhat similar in their goals, the standards were selected for evaluation by two separate groups with some common findings. STIX has had limited adoption by the financial sector but no single, definitive end user. The standard is still in development with the US government as the primary developer in partnership with OASIS. There is interest in using STIX to manage content, but users don't really care about what technology is used for the exchange. The initial goals may not wind up matching the end result for STIX, as managing content may be the primary use case. IODEF was specified by National Research and Education Networks (NRENs) and Computer Security Incident Response Teams (CSIRTs) and formalized in the IETF . The user is the security operations center (SOC). While there are several implementations, it is not widely adopted. In terms of exchange, users are more interested in indicators than full event information, and this applies to STIX as well. Sharing and trust are additional hurdles as many are not willing to disclose information. DNS-Based Authentication of Named Entities (DANE) has DNSSEC as a dependency, which is a hurdle toward adoption (too many dependencies). It has a roll-your-own adoption model, which is risky. While there are some large pockets of adoption, there is still much work to do to gain widespread adoption. A regulatory requirement gave rise to partial adoption in Germany, which naturally resulted in production of documentation written in German -- possibly giving rise to further adoption in German-speaking countries. There has also been progress made in the Netherlands through the creation of a website: . The website allows you to test your website for a number of standards (IPv6, DNSSEC, DANE, etc.). is a collaboration of industry organizations, companies, and the government in the Netherlands and is available for worldwide use. IP version 6 (IPv6) has struggled, and the expense of running a dual stack was one of the highest concerns on the list discussed in the workshop breakout. The end user for IPv6 is everyone, and the breakout team considered it too ambiguous. Too many new requirements have been added over its 20-year life. The scope of necessary adoption is large with many peripheral devices. Government requirements for support have helped somewhat with improved interoperability and adoption, but features like NAT being added to IPv4 slowed adoption. With no new features being added to IPv4 and lessons learned, there's still a possibility for success.

Breakout 2 Results: Preventative Protocols and Scaling Defense This breakout session followed the sessions on MUD, Protocol for Automated Vulnerability Assessment (PAVA), and Protocol for Automatic Security Configuration (PASC), which have themes of automation at scale. MUD was designed for Internet of Things (IoT), and as such, scaling was a major consideration. The PAVA and PASC work builds off of MUD and maintains some of the same themes. This breakout session was focused on groups brainstorming preventative measures and enabling vendors to deploy mitigations. One group dove a bit deeper into MUD and layer 2 (L2) discovery. MUD changes sets of filtering control management to the vendor or intermediary MUD vendors for a predictable platform that scales well. While the overall value of MUD is clear, the use of MUD and what traffic is expected for a particular device should be considered sensitive information, as it could be used to exploit a device. MUD has an option of using L2 discovery to share MUD files. L2 discovery, like the Dynamic Host Configuration Protocol (DHCP), is not encrypted from the local client to the DHCP server at this point in time (there is some interest to correct this, but it hasn't received enough support yet). As a result, it is possible to leak information and reveal data about the devices for which the MUD files would be applied. This could multicast out information such as network characteristics, firmware versions, manufacturers, etc. There was some discussion on the use of 802.11 to improve connections . Several participants from this group plan to research this further and identify options to prevent information leakage while achieving the stated goals of MUD. The next group discussed a proposal one of the participants had already begun developing, namely privacy for rendezvous service. The basic idea was to encrypt Server Name Indication (SNI) using DNS to obtain public keys. The suffix on server IPv6 would be unique to a TLS session (information missing). The discussion on this proposal was fruitful, as the full set of attendees engaged, with special interest from the incident responders to be involved in early review cycles. Incident responders are very interested to understand how protocols will change and to assess the overall impact of changes on privacy and security operations. Even if there are no changes to the protocol proposals stemming from this review, the group discussion landed on this being a valuable exchange to understand early the impacts of changes for incident detection and mitigation, to devise new strategies, and to provide assessments on the impact of protocol changes on security in the round. The third group reported back on trust exchanges relying heavily on relationships between individuals. They were concerned with scaling the trust model and finding ways to do that better. The group dove deeper into this topic. The fourth group discussed useful data for incident responders. This built on the first breakout session (). The group determined that indicators of compromise (IoCs) are what most organizations and groups are able to successfully exchange. Ideally, these would be fixed and programmable. They discussed developing a richer format for sharing event threats. When reporting back to the group, a successful solution used in the EU was mentioned: the Malware Information Sharing Platform (MISP). This will be considered in the review of existing efforts to determine if anything new is needed.

Breakout 3 Results: Incident Response Coordination Incident response coordination currently does not scale. This breakout session focused on brainstorming incident response and coordination, looking specifically at what works well for teams today, what is holding them back, and what risks loom ahead. Output from this session could be used to generate research and to dive deeper in a dedicated workshop on these topics. Supporting:

Trust between individuals in incident response teams
Volume of strong signals and automated discovery
Need to protect network as a forcing function
Law and legal catalyst, motivator to stay on top
Current efforts supported by profit and company interests, but those may shift
Fear initially results in activity or in terms of the diagram used, a burst of wind, but eventually leads to complacency

What creates drag:

Lack of clear Key Performance Indicators (KPIs)
Too many standards
Potential for regional borders to impact data flows
Ease of use for end users
Speed to market without security considerations
Legal framework slow to adapt
Disconnect in actual/perceived risk
Regulatory requirements preventing data sharing
Lack of clarity in shared information
Behind the problem/reactionary
Lack of resources/participation
Monoculture narrows focus

Looming problems:

Dynamic threat landscape
Liability
Vocabulary collision
Lack of target/adversary clarity
Bifurcation of Internet
Government regulation
Confusion around metrics
Sensitivity of intelligence (trust)
Lack of skilled analysts
Lack of "fraud loss" data sharing
Stakeholder/leader confusion
Unknown impact of emerging technologies
Overcentralization of the Internet
New technologies and protocols
Changes in application-layer configurations (e.g., browser resolvers)

Breakout 4 Results: Monitoring and Measurement The fourth breakout session followed 's talk on IPv6 aggregation to provide privacy for IPv6 sessions. Essentially, IPv6 provides additional capabilities for monitoring sessions end to end. Dave and his coauthor, , primarily focus on measurement research but found a way to aggregate sessions to assist with maintaining user privacy. If you can devise methods to perform management and measurement, or even perform security functions, while accommodating methods to protect privacy, a stronger result is likely. This also precludes the need for additional privacy improvement work to defeat measurement objectives. This breakout session was focused on devising methods to perform monitoring and measurement, coupled with advancing privacy considerations. The full group listed out options for protocols to explore and ranked them, with the four highest then explored by the breakout groups. Groups agreed to work further on the proposed ideas.

IP Address Reputation There is a need to understand address assignment and configuration for hosts and services, especially with IPv6 in (1) sharing IP-address-related information to inform attack response efforts while still protecting the privacy of victims and possible attackers and (2) mitigating abuse by altering the treatment, e.g., dropping or rate-limiting, of packets. Currently, there is no database that analysts and researchers can consult to, for instance, determine the lifetimes of IPv6 addresses or the prefix length at which the address is expected to be stable over time. The researchers propose either introducing a new database (compare PeeringDB) or extending existing databases (e.g., the regional Internet registries (RIRs)) to contain such information and allowing arbitrary queries. The prefix information would either be provided by networks that are willing or based on measurement algorithms that reverse-engineer reasonable values based on Internet measurements . In the former case, the incentive of networks to provide such information is to ensure that privacy of their users is respected and to limit collateral damage caused by access control lists affecting more of that network's addresses than necessary, e.g., in the face of abuse. This is an early idea; is the lead contact for those interested in helping to develop this further.

Server Name Authentication Reputation C (SNARC) SNARC is a mechanism to assign value to trust indicators, used to make decisions about good or bad actors. The mechanism would be able to distinguish between client and server connections and would be human readable. In addition, it builds on zero trust networking and avoids consolidation, thus supporting legitimate new players. SNARC has a similar theme to the IP reputation/BGP ranking idea mentioned above. SNARC is not currently defined by an RFC; however, such an RFC would help customers and design teams on existing solutions. The group plans to research visual aspects and underlying principles as they begin work on this idea. They plan to begin work in several stages, researching "trust" indicators, "trust" value calculations, and research actions to apply to "trust". The overarching goal is to address blind trust, one of the challenges identified with information/incident exchanges. is the lead contact for those interested in working with this team.

Logging The group presented the possibility of injecting logging capabilities at compile time for applications, resulting in a more consistent set of logs, covering an agreed set of conditions. Using a log-injecting compiler would increase logging for those applications and improve the uniformity of logged activity. Increasing logging capabilities at the endpoint is necessary as the shift toward increased use of encrypted transport continues. is the lead contact for those interested in developing this further.

Fingerprinting Fingerprinting has been used for numerous applications on the Web, including security, and will become of increasing importance with the deployment of stronger encryption. Fingerprinting provides a method to identify traffic without using decryption. The group discussed privacy considerations and balancing how you achieve the security benefits (identifying malicious traffic, information leakage, threat indicators, etc.). They are interested in deriving methods to validate the authenticity without identifying the source of traffic. They are also concerned with scaling issues. is the lead contact for those interested in working with this team.

Taxonomy and Gaps Session At the start of the second day of the workshop, and prepared (and Kirsty led) a workshop-style session to discuss taxonomies used in incident response, attacks, and threat detection, comparing solutions and identifying gaps. The primary objective was to determine a path forward by selecting the language to be used in the proposed SMART Research Group. Several taxonomies were presented for review and discussion. The topic remains open, but the following key points were highlighted by participants:

A single taxonomy might not be the way to go, because which taxonomy you use depends on what problem you are trying to solve, e.g., attribution of the attack, mitigation steps, technical features, or organizational impact measurements.
A tool to map between taxonomies should be automated, as there are requirements within groups or nations to use specific taxonomies.
The level of detail needed for reporting to management and for the analyst investigating the incident can be very different. At the workshop, one attendee mentioned that, for management reporting, they only use 8 categories to lighten the load on analysts, whereas some of the taxonomies contain 52 categories.
How you plan to use the taxonomy matters and may vary between use cases. Take, for instance, sharing data with external entities versus internal only. The taxonomy selected depends on what you plan to do with it. Some stated a need for attribute-based dynamic anthologies as opposed to rigid taxonomies used by others. A rigid taxonomy did not work for many from feedback in the session.
was briefly discussed as a possibility; however, there is a clear need to update terminology in this publication around this space in particular. This is likely to be raised in the Security Area Advisory Group (SAAG) during the open mic session, hopefully with proposed new definitions to demonstrate the issue and evolution of terms over time.
Within a taxonomy, prioritization matters to understand the impact of threats or an attack. How do you map that between differing taxonomies? What is the problem to be solved, and what tooling is required?
Attack attribution had varying degrees of interest. Some felt the public sector cared more about attribution, not about individuals. They were interested in possible motivations behind an attack and determining if there were other likely victims based on these motivations. Understanding if the source was an individual actor, organized crime, or a nation state mattered.

The result of this discussion was not to narrow down to one taxonomy but to think about mappings between taxonomies and the use cases for exchanging or sharing information, eventually giving rise to a common method to discuss threats and attacks. Researchers need a common vocabulary, not necessarily a common taxonomy.